100+ Gemini 3.5 Flash Prompts Cheat Sheet (2026)

Gemini 3.5 Flash costs roughly $0.30 per million input tokens. Claude Opus 4.7 costs $15. That is a 50× price gap — and for the vast majority of work you will ever throw at a language model, Flash gets you to the same answer in a fraction of the time at a fraction of the cost.

What is missing is rarely the model. It is the prompt.

This cheat sheet is 110 production-tested Gemini 3.5 Flash templates — pulled directly from the prompts we run at PromptsRush every day to ship product features, debug code, write copy, analyze data, extract from images, and orchestrate multi-step agents. Eleven categories. Zero recycled "10 ChatGPT prompts you have already seen" filler. Just the exact phrasings that hold up under real workload, ordered by what you are actually trying to accomplish.

Three numbers explain why Flash is the workhorse model to optimize for in 2026: a 1M-token context window that swallows an entire codebase, sub-400ms first-token latency that feels synchronous, and pricing low enough that you stop budgeting per call and start thinking in volume. The trade-off is real — Flash has a smaller reasoning head than Gemini Pro or Opus 4.7, and it rewards precision. Vague prompts come back vague. Every template below is engineered around that constraint.

If you want the strategic case for choosing Flash over the heavyweight models in the first place, our Gemini 3.5 Flash vs Claude Opus 4.7 for coding deep-dive is the companion piece. Read that one for the when. Read this one for the what to say.

How to Use This Cheat Sheet

Each prompt is a template. Square brackets like [LANGUAGE] are variables — replace them with your actual values before sending.
Long-press or select-and-copy on any prompt block to grab it. We deliberately avoided per-prompt copy buttons here because 110 of them would have made the page look like a slot machine.
Mix and match. Most real workflows chain 2–4 of these prompts together. The "Chaining" section near the end shows the patterns we use most.
Read the meta-rules first. They're what separates a Flash prompt that works from one that almost works.

Why Gemini 3.5 Flash Specifically

Three things make Flash the workhorse model of 2026:

1M-token context window — paste a whole repo, a whole book, a whole quarter of meeting notes, and ask questions across it.
Sub-400ms first-token latency in most regions — fast enough to feel synchronous.
~$0.30 input / ~$2.50 output per million tokens — cheap enough that you stop thinking about cost and start thinking about volume.

The flip side: Flash is a smaller reasoning head than the Gemini Pro tier or Claude Opus. It rewards specific, structured, output-constrained prompts. Vague asks come back vague. Every template below is built around that constraint.

10 Meta-Rules That Make Every Flash Prompt Better

Assign a role on line 1. "You are a senior data engineer" beats "act as a data engineer" because it's declarative, not performative.
Specify the output format up front. "Return JSON with keys X, Y, Z" before the actual task, not after.
Constrain length explicitly. "Under 80 words." "Exactly 3 bullets." "One paragraph, no headers." Flash will obey numeric constraints reliably.
Use delimiters for inputs. Wrap pasted content in ===INPUT=== / ===END=== markers. Flash is much better at separating instructions from content this way.
Ask for the reasoning second, the answer first. "Give me the answer, then a one-line rationale." Otherwise Flash buries the answer in preamble.
Use few-shot examples for anything stylistic. One good example beats five paragraphs of instructions.
Forbid hedging when you want a decision. "Pick one. No 'it depends.'"
For multi-step tasks, ask for a numbered plan first. Then run the steps as separate calls. Flash is much better at execution than open-ended synthesis.
Use the system prompt for persistent constraints. Style, tone, output format, banned words. Don't repeat them in every user turn.
Lower temperature for anything structured. temperature: 0.1 for code, JSON, SQL, data work. 0.7+ only for creative writing.

Pro tip: When a Flash prompt isn't working, the fix is almost never "add more instructions." It's usually "shorten it and add a concrete example."

1. Coding & Engineering (Prompts 1–10)

1. The Function Generator

Single-purpose utility functions, any language.

You are a senior [LANGUAGE] engineer. Write a function `[NAME]` that [DESCRIBE BEHAVIOR]. Requirements: (1) full type annotations, (2) handle [EDGE CASE 1] and [EDGE CASE 2], (3) include one usage example with expected output as a comment. Return code only, no prose.

2. The Bug Finder

For pasted code that's misbehaving.

You are a debugger. Code below. Identify the bug, explain it in one sentence, then output the corrected code. If there is no bug, say "NO BUG FOUND" and stop.
===CODE===
[PASTE CODE]
===END===

3. The Code Reviewer

PR-style review with actionable comments.

Review this [LANGUAGE] code as a senior engineer. Return a markdown list of issues, each formatted as: `- [SEVERITY] line X: issue → suggested fix`. Severities: BUG, PERF, STYLE, NIT. Skip nits unless the file is otherwise clean. Max 10 items.
===CODE===
[PASTE]
===END===

4. The Test Generator

Unit tests for a given function.

Write [PYTEST/JEST/VITEST] tests for the function below. Cover: happy path, empty input, malformed input, boundary values, and one obscure edge case. Use AAA structure (Arrange, Act, Assert). One assertion per test. Code only.
===FUNCTION===
[PASTE]
===END===

5. The Refactor Planner

Plan first, code second.

I want to refactor the code below to [GOAL]. Do not write any code yet. Produce a numbered plan of 3–7 atomic steps, each one a single conceptual change with a one-line rationale. After the plan, list any assumptions you're making.
===CODE===
[PASTE]
===END===

6. The Regex Builder

Regex you can actually read.

Write a [JAVASCRIPT/PYTHON/POSIX] regex that matches [DESCRIBE PATTERN]. Format your answer as: (1) the regex on its own line, (2) a one-sentence plain-English explanation, (3) three example strings it should match, (4) three example strings it should NOT match.

7. The Docstring Writer

Generate docs from code.

Add [JSDOC/NUMPY-STYLE/GOOGLE-STYLE] docstrings to every public function below. Document: purpose, parameters with types, return value, and one raise/throw condition. No marketing language. Code with docstrings only.
===CODE===
[PASTE]
===END===

8. The API Client Generator

Typed client from an OpenAPI snippet.

Given the OpenAPI endpoint below, generate a typed [TYPESCRIPT/PYTHON] client function. Include: request type, response type, fetch implementation with error handling for 4xx/5xx, and one usage example. No external dependencies unless I name them.
===SPEC===
[PASTE]
===END===

9. The SQL Writer

Postgres-flavored, explained.

You are a Postgres expert. Write a single SQL query that [DESCRIBE GOAL]. Schema below. Use CTEs for readability. After the query, add a one-line comment per CTE explaining what it does. No EXPLAIN, no setup.
===SCHEMA===
[PASTE TABLE DEFS]
===END===

10. The Commit Message Writer

Conventional-Commits-style messages from a diff.

Write a single conventional-commits message for the diff below. Format: `type(scope): subject` (≤72 chars) + blank line + 1–3 bullet body lines explaining WHY, not WHAT. Types: feat, fix, refactor, perf, docs, test, chore.
===DIFF===
[PASTE]
===END===

2. Data & Analytics (Prompts 11–20)

11. The CSV Summarizer

What's in this file?

Below is the first 20 rows of a CSV. Describe: (1) what each column likely represents, (2) data types, (3) any column that looks like it has dirty data, (4) three analytical questions this dataset could answer. Max 200 words.
===CSV===
[PASTE]
===END===

12. The SQL Query Optimizer

For slow queries.

Review this Postgres query for performance. Identify the 3 most likely bottlenecks and propose specific fixes (indexes, query rewrites, CTE replacements). Rank by expected impact. Do not rewrite the whole query — propose surgical changes only.
===QUERY===
[PASTE]
===EXPLAIN===
[PASTE EXPLAIN ANALYZE OUTPUT IF AVAILABLE]
===END===

13. The JSON-to-Schema Converter

JSON Schema or TS types from sample JSON.

Generate a [JSON-SCHEMA / TYPESCRIPT INTERFACE / ZOD SCHEMA] for the JSON below. Infer optional vs required by checking which keys appear in every object. For unions, use the narrowest type that fits. Output schema only, no commentary.
===JSON===
[PASTE]
===END===

14. The Data Validator

Find dirty rows fast.

Below is sample data. Define validation rules as a JSON list of `{field, rule, severity}` objects. Then identify any rows that violate these rules. Severities: ERROR (must fix), WARN (review). Output: rules JSON first, then violations table.
===DATA===
[PASTE]
===END===

15. The Anomaly Detector

Eyeball-grade outlier detection.

Time-series data below. Identify the top 5 anomalies. For each, output: timestamp, value, expected range, and a one-sentence hypothesis for the cause. Use simple statistical reasoning (rolling mean ± 3σ is fine). Format as a markdown table.
===DATA===
[PASTE]
===END===

16. The Pivot Summary

Quick aggregations without spinning up a notebook.

From the data below, produce a pivot summary showing [METRIC] by [DIMENSION 1] x [DIMENSION 2]. Include row totals, column totals, and a one-line "key insight" beneath. Output as a markdown table.
===DATA===
[PASTE]
===END===

17. The Plotting Prompt

Get a chart spec, not a chart.

Given the data below, recommend the single best chart type to visualize [QUESTION]. Then output a Vega-Lite spec for that chart. Justify the choice in one sentence above the spec. No matplotlib, no extra libraries.
===DATA===
[PASTE]
===END===

18. The Cohort Analyzer

Retention/usage cohorts from raw events.

Event log below. Compute weekly retention cohorts: rows = signup week, columns = weeks since signup, cells = % of cohort active that week. Format as a markdown table. Beneath, write 3 bullet observations about retention shape.
===EVENTS===
[PASTE]
===END===

19. The Funnel Analyzer

Conversion funnel from raw events.

Events below. Steps in order: [STEP 1] → [STEP 2] → [STEP 3] → [STEP 4]. For each step, output: users entering, % continuing, % dropping. After the table, identify the single biggest drop and propose 2 hypotheses for why.
===EVENTS===
[PASTE]
===END===

20. The Data Cleaner

Find-and-fix transformations for a dirty dataset.

Sample data below. Identify cleaning operations needed (dedupe, normalize, type-cast, fill nulls, etc.). Output a numbered list of transformations as plain English, then the equivalent Pandas code block. Operations should be idempotent.
===DATA===
[PASTE]
===END===

3. Writing & Editing (Prompts 21–30)

21. The Tone Matcher

Rewrite to match a style sample.

Rewrite the draft below in the voice of the sample. Match: sentence length, vocabulary range, use of contractions, and rhythm. Do not match: topic, structure. Output rewrite only.
===SAMPLE===
[PASTE 2–3 PARAGRAPHS OF SOURCE VOICE]
===DRAFT===
[PASTE DRAFT]
===END===

22. The Length Editor

Compress or expand to a target word count.

Rewrite the text below to exactly [N] words, ±5. Preserve every concrete fact, number, and proper noun. Cut adjectives, filler, hedging, and meta-commentary first. Output rewrite only.
===TEXT===
[PASTE]
===END===

23. The Clarity Rewriter

Make dense writing readable.

Rewrite the paragraph below for a non-expert reader at an 8th-grade reading level. Replace jargon with plain words. Break sentences over 25 words. Keep all technical accuracy. Same length, ±10%.
===PARAGRAPH===
[PASTE]
===END===

24. The Translator

Faithful, not literal.

Translate the text below from [SOURCE LANG] to [TARGET LANG]. Prioritize natural fluency in the target language over literal word-mapping. Preserve: proper nouns, numbers, technical terms. If a phrase has no idiomatic equivalent, translate the meaning and flag with [adapted: original phrase].
===TEXT===
[PASTE]
===END===

25. The Headline Generator

10 variants in 5 styles.

Generate 10 headline variants for an article about [TOPIC, KEY ANGLE]. Mix these styles, 2 each: (1) curiosity gap, (2) numbered list, (3) direct benefit, (4) contrarian, (5) "how to". Max 12 words per headline. Output as a numbered list with the style in brackets.

26. The Outline Builder

Article outlines that don't waste words.

Build an outline for an article titled "[TITLE]". Audience: [AUDIENCE]. Goal: [READER TAKEAWAY]. Output: hook (1 sentence), 6–10 H2 sections (each with a one-line "what this section argues"), conclusion (1 sentence). No fluff sections like "introduction" or "what is X" unless essential.

27. The Style Enforcer

Strunk & White / AP / your own style guide.

Edit the text below to comply with [STYLE GUIDE NAME OR PASTE RULES]. Make minimal changes — preserve voice and meaning. Output: edited text first, then a bullet list of the changes made (with rule names).
===TEXT===
[PASTE]
===END===

28. The Email Rewriter

Make sent emails sound less robotic.

Rewrite this email to be [WARMER / FIRMER / SHORTER / MORE DIRECT]. Keep all factual content. Subject line stays the same unless it's clearly bad. Max 1 ask per email. Output rewritten email only.
===EMAIL===
[PASTE]
===END===

29. The Long-Doc Summarizer

Stratified summary for long inputs.

Summarize the document below in three layers: (1) ONE-SENTENCE summary, (2) THREE-BULLET summary, (3) FIVE-PARAGRAPH summary covering setup, argument, evidence, counterpoint, conclusion. Each layer self-contained.
===DOC===
[PASTE]
===END===

30. The Quote Puller

Extract pull-quote candidates.

From the text below, extract the 5 most quotable sentences. Criteria: standalone meaning, vivid language, strong opinion, useful for social sharing. Output as a numbered list. For each, add a one-line "why this one" rationale.
===TEXT===
[PASTE]
===END===

4. Marketing & SEO (Prompts 31–40)

31. The Blog Post Outline

Search-intent-aware outlines.

Target keyword: "[KEYWORD]". Search intent: [INFORMATIONAL/COMMERCIAL/TRANSACTIONAL]. Build a blog post outline that beats the top 3 results. Output: title (≤60 chars), meta description (≤155 chars), 8–12 H2 sections, and 3 FAQ questions to target People-Also-Ask.

32. The Meta Tag Generator

Title + description + OG.

Page topic: "[PAGE TOPIC]". Primary keyword: [KEYWORD]. Generate: (1) 3 title tag variants (≤60 chars each, keyword in first 30), (2) 3 meta descriptions (≤155 chars, action verb, includes keyword), (3) one OG description (≤90 chars, hook-focused). Output as JSON.

33. The Twitter Thread Writer

Hook-driven, scannable threads.

Topic: [TOPIC]. Audience: [WHO]. Goal: [TAKEAWAY]. Write a 7-tweet thread. Tweet 1: hook (curiosity or contrarian, ≤220 chars, no emojis). Tweets 2–6: one idea each, ≤270 chars, can use line breaks. Tweet 7: CTA. Number each tweet.

34. The LinkedIn Post Writer

Personal-voice posts.

Topic: [TOPIC]. Tone: [REFLECTIVE/CONTRARIAN/CELEBRATORY]. Write a LinkedIn post: (1) first line is a 1-sentence hook that breaks pattern, (2) blank line, (3) 3–5 short paragraphs telling a story or making a point, (4) one-line closer with a question. Max 1300 chars. No hashtags.

The hardest 80 words to write.

Write the opening 80 words of a newsletter issue about [TOPIC]. Tone: [CASUAL/PROFESSIONAL]. Must include: a concrete hook (not "happy Monday"), a one-line preview of what's inside, and a reason to keep reading. Cut every word that doesn't earn its place.

36. The Hook Variation Generator

For ads, videos, headlines.

Product: [PRODUCT]. Audience pain point: [PAIN]. Generate 20 opening-line hooks in 4 styles, 5 each: (1) shocking stat, (2) story-in-one-sentence, (3) contrarian claim, (4) direct question. Max 15 words per hook. Output as a numbered list with style in brackets.

37. The Schema Markup Writer

Valid JSON-LD on request.

Generate JSON-LD schema markup for a [ARTICLE/PRODUCT/REVIEW/HOWTO/FAQ] page. Page details below. Output: valid JSON-LD inside a `<script type="application/ld+json">` tag. Use schema.org vocabulary. No extra commentary.
===PAGE DETAILS===
[FILL IN]
===END===

38. The Internal Link Suggester

Anchor text + target URL pairs.

New article below. Existing site URLs and their topics also below. Suggest 5–8 internal links to add to the new article. For each: anchor text (3–6 words, natural), target URL, and the exact sentence in the new article where it should be placed.
===NEW ARTICLE===
[PASTE]
===EXISTING URLS===
[PASTE LIST: url — topic]
===END===

39. The Keyword Cluster Builder

Topic clusters from a seed.

Seed keyword: "[KEYWORD]". Build a topic cluster: (1) one pillar topic, (2) 8–12 subtopics, each a candidate H2 or sibling article. For each subtopic, add: search intent type, estimated difficulty (low/med/high), and the kind of content that would win (guide, listicle, comparison, etc.).

40. The Competitor Page Analyzer

What to beat, what to ignore.

URL/content of a top-ranking competitor page below. Analyze: (1) what's good about it (3 bullets), (2) what's weak about it (3 bullets), (3) 5 specific things our version should do differently to win the rank. Be concrete — no "make it better."
===COMPETITOR===
[PASTE]
===END===

5. Productivity & Workflows (Prompts 41–50)

41. The Meeting Notes Distiller

Raw transcript below. Output: (1) one-paragraph summary, (2) decisions made (bulleted), (3) action items as `OWNER — TASK — DUE`, (4) open questions (bulleted), (5) one risk/blocker if any. Skip pleasantries. If a field is empty, write "—".
===TRANSCRIPT===
[PASTE]
===END===

42. The Calendar Planner

I have these tasks: [LIST WITH ROUGH DURATIONS AND DEADLINES]. My deep-work hours are [TIME RANGES]. Generate a tomorrow's calendar: 25-min Pomodoros with 5-min breaks, hardest task first, no more than 4 hours of deep work, buffer for meetings: [LIST MEETINGS]. Output as a timestamped schedule.

43. The Daily Standup Writer

From the activity log below, write a 3-line standup update: yesterday (what shipped), today (what I'm doing), blockers (or "none"). Each line ≤140 chars. No filler. No emojis.
===LOG===
[PASTE COMMITS / TASKS / NOTES]
===END===

44. The Email Triager

Below are the subject lines and senders of [N] unread emails. Sort them into 4 buckets: REPLY-TODAY, REPLY-THIS-WEEK, READ-ONLY, ARCHIVE. For each REPLY-TODAY item, draft a 1–2 sentence reply.
===INBOX===
[PASTE]
===END===

45. The Task Decomposer

Goal: "[GOAL]". Constraint: must be done by [DATE], by [WHO]. Decompose into 3–7 atomic tasks, each ≤2 hours of work, with clear "done when…" criteria. Output as a numbered list. Flag dependencies between tasks.

46. The Decision Matrix

I'm choosing between: [OPTION A], [OPTION B], [OPTION C]. Criteria that matter to me: [LIST CRITERIA WITH WEIGHTS 1–5]. Build a weighted scoring table. Score each option 1–5 per criterion. Output: markdown table, weighted total, one-line recommendation.

47. The Status Report Generator

This week's activity below. Write a stakeholder-facing weekly status report: (1) headline (1 sentence), (2) shipped this week (bullets), (3) in flight (bullets), (4) risks / asks (bullets, or "none"), (5) next week's focus (bullets). Total ≤200 words.
===ACTIVITY===
[PASTE]
===END===

48. The OKR Drafter

Theme: [QUARTERLY THEME]. Team: [TEAM]. Draft 1 objective + 3 key results. Objective must be qualitative and inspirational. KRs must be measurable, time-bound, and have a baseline + target. No vanity metrics. Output formatted as `O: ...` and `KR1/2/3: from X to Y by [DATE]`.

49. The Weekly Reviewer

Below are this week's wins, misses, and blockers. Run a structured weekly review: (1) what worked (2–3 bullets), (2) what didn't (2–3 bullets), (3) one lesson, (4) one thing to do differently next week, (5) one thing to stop doing entirely.
===INPUT===
[PASTE]
===END===

50. The Inbox Zero Coach

I have [N] unread emails. Help me get to zero in 30 minutes. Output a 5-step plan: (1) which categories to bulk-archive first, (2) which folder rules to set up, (3) a script for the 3 email templates I'll need most, (4) a 60-second rule for the rest, (5) what to do tomorrow to keep it at zero.

6. Customer Support & Sales (Prompts 51–60)

51. The Empathetic Support Reply

Customer message below. Reply: (1) acknowledge the specific frustration in their own words, (2) explain what happened in plain language, (3) state what you're doing about it with a timeline, (4) end with a single concrete next step. No "we apologize for any inconvenience." Max 150 words.
===MESSAGE===
[PASTE]
===END===

52. The Firm Support Reply

Customer is asking for [REQUEST] that violates [POLICY]. Reply: (1) one-line empathetic acknowledgment, (2) clear no with the specific policy reason, (3) one alternative we CAN offer, (4) close politely. Do not apologize for the policy. Max 120 words.
===MESSAGE===
[PASTE]
===END===

53. The Refund Decision

Refund request below. Decide: APPROVE, PARTIAL, or DENY. Reasoning in one paragraph, citing: customer history, request validity, business impact. Then draft the customer-facing reply matching the decision. Output: decision, reasoning, reply.
===REQUEST===
[PASTE]
===CUSTOMER HISTORY===
[PASTE]
===END===

54. The Knowledge Base Article

Question we get often: "[QUESTION]". Write a KB article: (1) H1 phrased as the user's question, (2) one-paragraph answer at the top, (3) step-by-step instructions if applicable, (4) "When this doesn't work" troubleshooting section, (5) "Related" section with 2–3 link suggestions. Max 400 words.

55. The Sales Follow-Up

Last contact with prospect: [SUMMARIZE WHAT WAS DISCUSSED, WHEN]. Their stated objection or hesitation: [OBJECTION]. Write a follow-up email that: (1) references something specific from last conversation, (2) addresses the objection with one concrete proof point, (3) proposes a low-friction next step. Max 110 words.

56. The Cold Outreach

Prospect: [NAME, TITLE, COMPANY]. Their company's likely pain point: [PAIN]. Our offer: [OFFER]. Write a cold email: (1) personalized opener referencing something real about them or their company, (2) one-sentence relevance bridge, (3) one specific value claim with a number, (4) low-friction ask (15-min call OR just a reply). Max 90 words. No "I hope this finds you well."

57. The Discovery Question Generator

I'm about to do a discovery call with [TITLE] at a [COMPANY TYPE]. Goal: understand if our [PRODUCT] fits. Generate 12 questions in 4 buckets, 3 each: (1) current state, (2) goals, (3) gaps/pain, (4) decision process. Order them so the call flows naturally. Each question ≤20 words.

58. The Objection Handler

Customer just said: "[OBJECTION]". I'm selling [PRODUCT/SERVICE]. Generate 3 ways to respond: (1) acknowledge + reframe, (2) acknowledge + counter with proof, (3) acknowledge + question back to surface the real concern. Each response ≤60 words. Don't be defensive.

59. The Demo Script

Product: [PRODUCT]. Audience: [TITLE]. Their use case: [USE CASE]. Their top concern: [CONCERN]. Build a 15-minute demo script: (1) 1-min opening (mirror back their problem), (2) 10 min of demo split into 3 chapters with what to show and what to say, (3) 2-min question prompt, (4) 2-min next-steps close.

60. The Renewal Email

Customer renewal in [N] days. Their usage this year: [USAGE METRICS]. Their original goal when signing: [GOAL]. Write a renewal email: (1) lead with a specific outcome they got, (2) name one feature they've underused that could compound the value, (3) renewal CTA. No discounts unless I say so. Max 130 words.

7. Research & Summarization (Prompts 61–70)

61. The Paper TL;DR

Academic paper below. Output: (1) one-sentence TL;DR, (2) the problem in plain language, (3) the method in 2 sentences, (4) the result with the headline number, (5) one caveat, (6) one practical implication. Total ≤180 words.
===PAPER===
[PASTE ABSTRACT + INTRO + CONCLUSION]
===END===

62. The Technical Paper Summary

Paper below. Summarize for a [DOMAIN] practitioner. Cover: prior work it builds on, what's actually new, experimental setup, key results with numbers, threats to validity, where it would and wouldn't generalize. 400–500 words. Markdown headers.
===PAPER===
[PASTE]
===END===

63. The Multi-Doc Synthesizer

Below are [N] sources on [TOPIC]. Produce: (1) the points of consensus across sources (bulleted), (2) the points of disagreement (with which source takes which side), (3) the strongest single claim from each source, (4) one open question none of them answer.
===SOURCES===
[PASTE WITH `SOURCE 1: ...` HEADERS]
===END===

64. The Quote Extractor

From the source below, extract every direct quote attributable to [PERSON/ROLE]. For each: the exact quote (verbatim), context (1 sentence), and page/timestamp if available. Output as a markdown table.
===SOURCE===
[PASTE]
===END===

65. The Counter-Argument Finder

Argument: "[CLAIM]". Steelman the 3 strongest counter-arguments. For each: (1) the counter-claim in one sentence, (2) the strongest evidence supporting it, (3) the kind of person who would find it most persuasive. Be fair to the counter side.

66. The Source Credibility Check

Source below. Rate credibility on 5 dimensions, each 1–5: author expertise, evidence quality, transparency about methods, recency, bias indicators. Provide a one-line justification per dimension. End with an overall verdict: HIGH / MEDIUM / LOW / DO-NOT-CITE.
===SOURCE===
[PASTE URL + KEY CONTENT]
===END===

67. The Glossary Builder

Domain: [DOMAIN]. Document below. Extract every domain-specific term used. For each: term, plain-English definition (≤25 words), and the sentence from the doc where it's used. Sort alphabetically. Output as a markdown table.
===DOC===
[PASTE]
===END===

68. The Timeline Builder

Topic: [TOPIC]. Sources below. Build a chronological timeline of key events. For each event: date (as precise as possible), 1-sentence what happened, source citation. Mark uncertain dates with [~]. Output: markdown table sorted oldest-first.
===SOURCES===
[PASTE]
===END===

69. The Comparison Table

Compare [ITEMS]: [LIST]. Dimensions to compare: [LIST]. For each cell, give a 5–12 word answer based ONLY on the sources below. If a source doesn't cover something, write "—" rather than inventing. Output: markdown table. After the table, name a winner per dimension.
===SOURCES===
[PASTE]
===END===

70. The Briefing Memo

Audience: [DECISION MAKER ROLE]. They have 5 minutes. Topic: [TOPIC]. Sources below. Write a briefing memo: (1) BLUF (Bottom Line Up Front, 2 sentences), (2) background (3 bullets), (3) options (2–4, each with pros/cons), (4) recommendation with rationale, (5) what we need to decide today. Max 400 words.
===SOURCES===
[PASTE]
===END===

8. Learning & Teaching (Prompts 71–80)

71. The ELI5 Explainer

Explain [CONCEPT] to a curious 12-year-old who is good at math but has no domain knowledge. Use one extended analogy that runs through the whole explanation. End with one "and that's why it matters" sentence. Max 200 words. No jargon.

72. The Socratic Tutor

I'm learning [TOPIC] and I think I understand [SPECIFIC POINT]. Don't tell me whether I'm right. Instead, ask me 3 questions that will reveal where my understanding breaks down if it does. Wait for my answers before continuing.

73. The Quiz Generator

Source material below. Generate 10 multiple-choice questions covering its key ideas. 4 options each, 1 correct. Mix difficulty: 4 recall, 4 application, 2 synthesis. Output as JSON: `[{question, options:[], correctIndex, explanation}, ...]`.
===SOURCE===
[PASTE]
===END===

74. The Flashcard Generator

From the source below, generate flashcards in Anki cloze-deletion format. 15 cards. Each card: one fact, one cloze, ≤25 words. Test the most testable, not the most interesting. Output one card per line as `Card N: text with {{c1::cloze}} deleted`.
===SOURCE===
[PASTE]
===END===

75. The Lesson Plan

Topic: [TOPIC]. Audience: [LEARNER LEVEL]. Time: [N] minutes. Build a lesson plan: (1) learning objective (1 sentence), (2) prerequisite knowledge check (3 questions), (3) explanation phase with one example, (4) practice phase with 2 exercises, (5) check-for-understanding (1 question). Include rough timing.

76. The Curriculum Gap Finder

Below is what I've already learned about [TOPIC]. Identify the 5 most important gaps in my knowledge — concepts I'm missing that would block deeper understanding. For each: what it is (1 sentence), why it matters, and the single best resource to learn it (book, paper, tutorial).
===WHAT I KNOW===
[PASTE]
===END===

77. The Code-Along Walkthrough

Teach me [CONCEPT] in [LANGUAGE] by walking me through building a tiny working example. Format: alternating between a code snippet (≤10 lines) and a 2-sentence explanation of what just happened and why. Build incrementally. End with the full working file.

78. The Mental Model Explainer

I want a mental model for [TOPIC]. Give me one — exactly one — model that captures the most important dynamic. Output: (1) name the model, (2) describe it in 100 words, (3) one concrete example of it in action, (4) one place the model breaks down. No "there are many ways to think about this."

79. The Worked Example Generator

Concept: [CONCEPT]. Generate a worked example: (1) state the problem precisely, (2) solve it step by step, with the reasoning for each step in italics, (3) state the final answer, (4) add one "common mistake to avoid" callout. Domain: [MATH/CODE/REASONING/ETC].

80. The Reading List Builder

Goal: become competent in [TOPIC] in [TIMEFRAME]. Current level: [LEVEL]. Build a sequenced reading list of 8–12 items. For each: title, author, format (book/paper/post), estimated time, why it's on the list at that position. Order matters — earlier items should unlock later ones.

9. Creative & Brainstorming (Prompts 81–90)

81. The Naming Generator

I'm naming a [PRODUCT/FEATURE/COMPANY/CHARACTER]. It does [DESCRIPTION]. The vibe should be [VIBE]. Generate 20 names in 4 styles, 5 each: (1) made-up words, (2) compound nouns, (3) metaphors, (4) verbs. For each: the name and a one-sentence rationale. Avoid AI cliches (Nexus, Synergy, Catalyst, etc.).

82. The Tagline Generator

Product: [PRODUCT]. Core promise: [PROMISE]. Audience: [WHO]. Generate 15 tagline options across these registers: 5 declarative, 5 imperative, 5 questioning. Max 8 words each. No "imagine if…" or "the future of…" openers.

83. The Story Premise Generator

Genre: [GENRE]. Constraint: [CONSTRAINT — e.g., must take place in one room, must have an unreliable narrator]. Generate 5 story premises. Each: 1 logline (≤25 words), the central conflict in one sentence, the moment when things change.

84. The Title Page Test

Below is the opening 500 words of a piece of writing. If a reader saw only this, would they keep reading? Rate: hook strength (1–10), clarity of stakes (1–10), voice distinctiveness (1–10). Then identify the single weakest sentence and propose a stronger replacement.
===OPENING===
[PASTE]
===END===

85. The SCAMPER Ideator

Subject: [PRODUCT/PROCESS/CONCEPT]. Apply SCAMPER. For each letter (Substitute, Combine, Adapt, Modify, Put to other uses, Eliminate, Reverse), generate 2 ideas. Each idea: one sentence describing the change, one sentence on why it could matter. Skip ideas that are obvious or trivial.

86. The Lateral Thinker

Problem: [PROBLEM]. Most obvious solutions: [LIST 2–3 IF YOU HAVE THEM]. Generate 5 lateral solutions — ideas that come from a completely different domain or first principle. Each: the idea (1 sentence), the domain it borrows from, what would have to be true for it to work.

87. The Crazy 8s Prompt

Design challenge: [CHALLENGE]. Constraint: [ONE HARD CONSTRAINT]. Generate 8 wildly different concept directions in 1 round. Each: name, one-sentence description, one-line rationale. Range from safest to weirdest. Don't self-censor — bad ideas welcome.

88. The Pitch Deck Outliner

Company: [STARTUP]. Audience: [SEED / SERIES A / GROWTH INVESTORS]. Build a 10-slide pitch deck outline. For each slide: title, one-line of what to argue, and the single most important number or visual to include. Story arc must build to the ask.

89. The World-Building Prompt

Setting: [TIME, PLACE, GENRE]. Generate a world bible: (1) one defining physical fact about this world, (2) one social rule everyone follows, (3) one technology or magic system rule, (4) the central tension, (5) the kind of conflict that drives stories here. Each in 2–3 sentences.

90. The Dialogue Writer

Scene: [SCENE]. Characters: [A] who wants [GOAL A], [B] who wants [GOAL B]. They cannot both win. Write the scene as dialogue only, no action lines. Subtext > text. End on a line that shifts the power balance.

10. Multimodal: Vision & Audio (Prompts 91–100)

These leverage Flash's strong multimodal input — images, screenshots, charts, audio. Paste the prompt and attach the file.

91. The Image Describer

Describe the attached image in three layers: (1) one-sentence summary, (2) factual contents (objects, people, setting, text visible), (3) interpretive layer (mood, likely context, audience). Total ≤200 words. Flag anything ambiguous with [uncertain: ...].

92. The Diagram-to-Code

The attached image is a [DIAGRAM TYPE: flowchart / ERD / sequence diagram / state machine]. Convert it to [MERMAID / PLANTUML / DOT] source. Preserve every node, edge, label, and direction. Output code block only.

93. The Whiteboard-to-Doc

The attached image is a photo of a whiteboard from a meeting on [TOPIC]. Convert to a clean markdown document: extract every bullet, arrow, and diagram into prose + structured lists. Resolve handwriting ambiguities with [?]. Note which sections seemed most emphasized (boxed, underlined, starred).

94. The Receipt-to-JSON

Attached image is a receipt. Extract: vendor name, date, line items (name, qty, unit price, total), subtotal, tax, total, payment method, last 4 of card. Output as JSON. If a field is illegible, use null and add a `_warnings` array listing which fields were uncertain.

95. The Chart-to-Summary

The attached image is a chart. Output: (1) chart type, (2) what it's measuring (axes, units), (3) the headline finding in one sentence, (4) the most surprising data point, (5) one question the chart raises but doesn't answer. Do not invent precise numbers — describe shapes and rough magnitudes only.

96. The Screenshot-to-Bug-Report

Attached screenshot shows a UI issue I'm reporting. Write a bug report: (1) summary (1 sentence), (2) steps to reproduce (inferred + ask me for what's missing), (3) expected behavior, (4) actual behavior visible in screenshot, (5) severity guess with rationale. Output as markdown.

97. The Slide Deck Extractor

Attached image is a slide from a presentation. Extract: title, subtitle if any, every bullet/text element verbatim, presenter notes if visible, and a 1-sentence guess at the slide's purpose in a deck. Preserve the visual hierarchy in the output.

98. The Audio Transcript Cleaner

Raw transcript below (auto-generated, messy). Clean it: remove filler words (um, like, you know), fix obvious mis-hearings using context, add paragraph breaks at topic shifts, add speaker labels if inferable. Do NOT paraphrase — keep exact wording where intelligible. Output cleaned transcript only.
===TRANSCRIPT===
[PASTE]
===END===

99. The Audio-to-Action-Items

Attached audio is a [MEETING / VOICE NOTE / CALL]. Extract action items only. Format: `OWNER — TASK — DUE`. If owner or due date is implicit, infer with [inferred]. Skip discussion that didn't lead to a decision. Order by mentioned-priority. Max 15 items.

100. The Video Frame Critique

Attached frame is from a [SHORT-FORM VIDEO / AD / TUTORIAL] at the [HOOK / MID / OUTRO] position. Critique it for that role: (1) what's working, (2) what's weak, (3) one specific change that would improve performance. Be concrete — no "make it more engaging."

11. Bonus: Agentic & Tool-Use (Prompts 101–110)

These are the prompts we use when Flash is one model in a larger agent loop — orchestrated by something like a planner-model on Claude Opus or a workflow tool like Genspark.

101. The Plan-Then-Execute Agent

You are an agent. Goal: [GOAL]. Available tools: [LIST]. Output a numbered plan first, max 7 steps. Wait for me to type "GO" before executing. When executing, do one step per response, output: which step, what tool you'll call, why, and what you expect back. After the result, decide: continue / revise plan / done.

102. The Tool Router

You are a routing layer. Given a user message and the tool registry below, output ONLY a JSON object: `{"tool": "tool_name", "args": {...}, "confidence": 0.0-1.0, "reasoning": "one sentence"}`. If no tool fits, return `{"tool": null, "reasoning": "..."}`. Never call the tool yourself.
===TOOLS===
[PASTE TOOL SCHEMAS]
===USER MESSAGE===
[PASTE]
===END===

103. The Self-Critique Loop

Below is a draft answer to "[ORIGINAL QUESTION]". Critique it as a skeptical expert: (1) 3 specific weaknesses, (2) 1 factual claim that should be verified, (3) 1 missing perspective. Then rewrite the answer addressing every critique. Output: critique first, then rewrite.
===DRAFT===
[PASTE]
===END===

104. The JSON-Only Enforcer

You are a JSON API. The user will describe what they want. You will reply with valid JSON only — no markdown, no code fences, no preamble, no explanation. If the request is impossible or ambiguous, return `{"error": "...reason..."}`. Schema for valid responses: [SCHEMA]. User request follows.

105. The Function Call Schema Builder

Function name: [NAME]. What it does: [DESCRIPTION]. Generate a function-calling schema in [OPENAI / ANTHROPIC / GEMINI] format. Include: every required parameter with type and description, every optional parameter with default, return type description, and one example call.

106. The Multi-Agent Orchestrator

Task: [TASK]. Available agents: [LIST: name → specialty]. Decompose the task and assign sub-tasks. Output as JSON: `{"workflow":[{"agent":"X","task":"...","depends_on":[]}], "estimated_steps": N}`. Make dependencies explicit. Do not execute — just plan.

107. The RAG Prompt

You are answering using ONLY the context below. Rules: (1) if the answer isn't in the context, say "I don't know based on the provided sources" — do not use outside knowledge, (2) cite the source ID for every claim like [src:3], (3) if sources disagree, say so and present both sides.
===QUESTION===
[USER QUESTION]
===CONTEXT===
[RETRIEVED CHUNKS WITH IDs]
===END===

108. The Long-Context Loader

I'm about to paste a [DOCUMENT TYPE] of ~[N] tokens. First, acknowledge the upload with the word "LOADED" and nothing else. Then I'll ask questions about it. When answering, always cite the section header or page number you're drawing from. Treat sections not in the doc as "not addressed."

109. The Cost-Aware Router

Task description: "[TASK]". Available models: Flash (cheap, fast, weaker reasoning), Pro (mid, balanced), Opus (expensive, strongest reasoning). Recommend the cheapest model that's likely to succeed. Output: model name, confidence (0–1), reasoning (1 sentence), and what to escalate to if the cheaper model fails.

110. The Confidence Rater

Below is my draft answer to "[ORIGINAL QUESTION]". Rate your confidence in each factual claim from 0–100. Output as a markdown table: claim | confidence | what would change your confidence. Then give the overall confidence in the answer as a whole and identify the single claim most worth verifying first.

How to Chain These Into Pipelines

Single prompts get you single answers. Chains get you systems. Five patterns we use weekly:

Plan → Execute → Critique: Use Prompt 5 (Refactor Planner) → Prompts 1–4 (Coding) → Prompt 103 (Self-Critique). Three model calls, dramatically better output than any single call.
Research → Synthesize → Brief: Prompt 63 (Multi-Doc Synthesizer) → Prompt 65 (Counter-Argument) → Prompt 70 (Briefing Memo). The decision memos we send executives all run through this exact chain.
Extract → Validate → Load: Prompt 94 (Receipt-to-JSON) → Prompt 14 (Data Validator) → your ETL pipeline. Flash is excellent as the extraction layer in a larger data system.
Cold outreach scale-up: Prompt 56 (Cold Outreach) running per-prospect → Prompt 27 (Style Enforcer) for brand-voice consistency → Prompt 110 (Confidence Rater) to gate which ones get sent vs which need a human eye.
Long-doc Q&A: Prompt 108 (Long-Context Loader) → Prompt 107 (RAG Prompt) with the doc as context → Prompt 110 (Confidence Rater) on every answer. This is most enterprise "chat with your docs" features in 100 tokens.

What is missing is rarely the model. It is the prompt.

How to Use This Cheat Sheet

Each prompt is a template. Square brackets like [LANGUAGE] are variables — replace them with your actual values before sending.
Long-press or select-and-copy on any prompt block to grab it. We deliberately avoided per-prompt copy buttons here because 110 of them would have made the page look like a slot machine.
Mix and match. Most real workflows chain 2–4 of these prompts together. The "Chaining" section near the end shows the patterns we use most.
Read the meta-rules first. They're what separates a Flash prompt that works from one that almost works.

Why Gemini 3.5 Flash Specifically

Three things make Flash the workhorse model of 2026:

1M-token context window — paste a whole repo, a whole book, a whole quarter of meeting notes, and ask questions across it.
Sub-400ms first-token latency in most regions — fast enough to feel synchronous.
~$0.30 input / ~$2.50 output per million tokens — cheap enough that you stop thinking about cost and start thinking about volume.

10 Meta-Rules That Make Every Flash Prompt Better

Assign a role on line 1. "You are a senior data engineer" beats "act as a data engineer" because it's declarative, not performative.
Specify the output format up front. "Return JSON with keys X, Y, Z" before the actual task, not after.
Constrain length explicitly. "Under 80 words." "Exactly 3 bullets." "One paragraph, no headers." Flash will obey numeric constraints reliably.
Use delimiters for inputs. Wrap pasted content in ===INPUT=== / ===END=== markers. Flash is much better at separating instructions from content this way.
Ask for the reasoning second, the answer first. "Give me the answer, then a one-line rationale." Otherwise Flash buries the answer in preamble.
Use few-shot examples for anything stylistic. One good example beats five paragraphs of instructions.
Forbid hedging when you want a decision. "Pick one. No 'it depends.'"
For multi-step tasks, ask for a numbered plan first. Then run the steps as separate calls. Flash is much better at execution than open-ended synthesis.
Use the system prompt for persistent constraints. Style, tone, output format, banned words. Don't repeat them in every user turn.
Lower temperature for anything structured. temperature: 0.1 for code, JSON, SQL, data work. 0.7+ only for creative writing.

Pro tip: When a Flash prompt isn't working, the fix is almost never "add more instructions." It's usually "shorten it and add a concrete example."

1. Coding & Engineering (Prompts 1–10)

1. The Function Generator

Single-purpose utility functions, any language.

You are a senior [LANGUAGE] engineer. Write a function `[NAME]` that [DESCRIBE BEHAVIOR]. Requirements: (1) full type annotations, (2) handle [EDGE CASE 1] and [EDGE CASE 2], (3) include one usage example with expected output as a comment. Return code only, no prose.

2. The Bug Finder

For pasted code that's misbehaving.

You are a debugger. Code below. Identify the bug, explain it in one sentence, then output the corrected code. If there is no bug, say "NO BUG FOUND" and stop.
===CODE===
[PASTE CODE]
===END===

3. The Code Reviewer

PR-style review with actionable comments.

Review this [LANGUAGE] code as a senior engineer. Return a markdown list of issues, each formatted as: `- [SEVERITY] line X: issue → suggested fix`. Severities: BUG, PERF, STYLE, NIT. Skip nits unless the file is otherwise clean. Max 10 items.
===CODE===
[PASTE]
===END===

4. The Test Generator

Unit tests for a given function.

Write [PYTEST/JEST/VITEST] tests for the function below. Cover: happy path, empty input, malformed input, boundary values, and one obscure edge case. Use AAA structure (Arrange, Act, Assert). One assertion per test. Code only.
===FUNCTION===
[PASTE]
===END===

5. The Refactor Planner

Plan first, code second.

I want to refactor the code below to [GOAL]. Do not write any code yet. Produce a numbered plan of 3–7 atomic steps, each one a single conceptual change with a one-line rationale. After the plan, list any assumptions you're making.
===CODE===
[PASTE]
===END===

6. The Regex Builder

Regex you can actually read.

Write a [JAVASCRIPT/PYTHON/POSIX] regex that matches [DESCRIBE PATTERN]. Format your answer as: (1) the regex on its own line, (2) a one-sentence plain-English explanation, (3) three example strings it should match, (4) three example strings it should NOT match.

7. The Docstring Writer

Generate docs from code.

Add [JSDOC/NUMPY-STYLE/GOOGLE-STYLE] docstrings to every public function below. Document: purpose, parameters with types, return value, and one raise/throw condition. No marketing language. Code with docstrings only.
===CODE===
[PASTE]
===END===

8. The API Client Generator

Typed client from an OpenAPI snippet.

Given the OpenAPI endpoint below, generate a typed [TYPESCRIPT/PYTHON] client function. Include: request type, response type, fetch implementation with error handling for 4xx/5xx, and one usage example. No external dependencies unless I name them.
===SPEC===
[PASTE]
===END===

9. The SQL Writer

Postgres-flavored, explained.

You are a Postgres expert. Write a single SQL query that [DESCRIBE GOAL]. Schema below. Use CTEs for readability. After the query, add a one-line comment per CTE explaining what it does. No EXPLAIN, no setup.
===SCHEMA===
[PASTE TABLE DEFS]
===END===

10. The Commit Message Writer

Conventional-Commits-style messages from a diff.

Write a single conventional-commits message for the diff below. Format: `type(scope): subject` (≤72 chars) + blank line + 1–3 bullet body lines explaining WHY, not WHAT. Types: feat, fix, refactor, perf, docs, test, chore.
===DIFF===
[PASTE]
===END===

2. Data & Analytics (Prompts 11–20)

11. The CSV Summarizer

What's in this file?

Below is the first 20 rows of a CSV. Describe: (1) what each column likely represents, (2) data types, (3) any column that looks like it has dirty data, (4) three analytical questions this dataset could answer. Max 200 words.
===CSV===
[PASTE]
===END===

12. The SQL Query Optimizer

For slow queries.

Review this Postgres query for performance. Identify the 3 most likely bottlenecks and propose specific fixes (indexes, query rewrites, CTE replacements). Rank by expected impact. Do not rewrite the whole query — propose surgical changes only.
===QUERY===
[PASTE]
===EXPLAIN===
[PASTE EXPLAIN ANALYZE OUTPUT IF AVAILABLE]
===END===

13. The JSON-to-Schema Converter

JSON Schema or TS types from sample JSON.

Generate a [JSON-SCHEMA / TYPESCRIPT INTERFACE / ZOD SCHEMA] for the JSON below. Infer optional vs required by checking which keys appear in every object. For unions, use the narrowest type that fits. Output schema only, no commentary.
===JSON===
[PASTE]
===END===

14. The Data Validator

Find dirty rows fast.

Below is sample data. Define validation rules as a JSON list of `{field, rule, severity}` objects. Then identify any rows that violate these rules. Severities: ERROR (must fix), WARN (review). Output: rules JSON first, then violations table.
===DATA===
[PASTE]
===END===

15. The Anomaly Detector

Eyeball-grade outlier detection.

Time-series data below. Identify the top 5 anomalies. For each, output: timestamp, value, expected range, and a one-sentence hypothesis for the cause. Use simple statistical reasoning (rolling mean ± 3σ is fine). Format as a markdown table.
===DATA===
[PASTE]
===END===

16. The Pivot Summary

Quick aggregations without spinning up a notebook.

From the data below, produce a pivot summary showing [METRIC] by [DIMENSION 1] x [DIMENSION 2]. Include row totals, column totals, and a one-line "key insight" beneath. Output as a markdown table.
===DATA===
[PASTE]
===END===

17. The Plotting Prompt

Get a chart spec, not a chart.

Given the data below, recommend the single best chart type to visualize [QUESTION]. Then output a Vega-Lite spec for that chart. Justify the choice in one sentence above the spec. No matplotlib, no extra libraries.
===DATA===
[PASTE]
===END===

18. The Cohort Analyzer

Retention/usage cohorts from raw events.

Event log below. Compute weekly retention cohorts: rows = signup week, columns = weeks since signup, cells = % of cohort active that week. Format as a markdown table. Beneath, write 3 bullet observations about retention shape.
===EVENTS===
[PASTE]
===END===

19. The Funnel Analyzer

Conversion funnel from raw events.

Events below. Steps in order: [STEP 1] → [STEP 2] → [STEP 3] → [STEP 4]. For each step, output: users entering, % continuing, % dropping. After the table, identify the single biggest drop and propose 2 hypotheses for why.
===EVENTS===
[PASTE]
===END===

20. The Data Cleaner

Find-and-fix transformations for a dirty dataset.

Sample data below. Identify cleaning operations needed (dedupe, normalize, type-cast, fill nulls, etc.). Output a numbered list of transformations as plain English, then the equivalent Pandas code block. Operations should be idempotent.
===DATA===
[PASTE]
===END===

3. Writing & Editing (Prompts 21–30)

21. The Tone Matcher

Rewrite to match a style sample.

Rewrite the draft below in the voice of the sample. Match: sentence length, vocabulary range, use of contractions, and rhythm. Do not match: topic, structure. Output rewrite only.
===SAMPLE===
[PASTE 2–3 PARAGRAPHS OF SOURCE VOICE]
===DRAFT===
[PASTE DRAFT]
===END===

22. The Length Editor

Compress or expand to a target word count.

Rewrite the text below to exactly [N] words, ±5. Preserve every concrete fact, number, and proper noun. Cut adjectives, filler, hedging, and meta-commentary first. Output rewrite only.
===TEXT===
[PASTE]
===END===

23. The Clarity Rewriter

Make dense writing readable.

Rewrite the paragraph below for a non-expert reader at an 8th-grade reading level. Replace jargon with plain words. Break sentences over 25 words. Keep all technical accuracy. Same length, ±10%.
===PARAGRAPH===
[PASTE]
===END===

24. The Translator

Faithful, not literal.

Translate the text below from [SOURCE LANG] to [TARGET LANG]. Prioritize natural fluency in the target language over literal word-mapping. Preserve: proper nouns, numbers, technical terms. If a phrase has no idiomatic equivalent, translate the meaning and flag with [adapted: original phrase].
===TEXT===
[PASTE]
===END===

25. The Headline Generator

10 variants in 5 styles.

Generate 10 headline variants for an article about [TOPIC, KEY ANGLE]. Mix these styles, 2 each: (1) curiosity gap, (2) numbered list, (3) direct benefit, (4) contrarian, (5) "how to". Max 12 words per headline. Output as a numbered list with the style in brackets.

26. The Outline Builder

Article outlines that don't waste words.

Build an outline for an article titled "[TITLE]". Audience: [AUDIENCE]. Goal: [READER TAKEAWAY]. Output: hook (1 sentence), 6–10 H2 sections (each with a one-line "what this section argues"), conclusion (1 sentence). No fluff sections like "introduction" or "what is X" unless essential.

27. The Style Enforcer

Strunk & White / AP / your own style guide.

Edit the text below to comply with [STYLE GUIDE NAME OR PASTE RULES]. Make minimal changes — preserve voice and meaning. Output: edited text first, then a bullet list of the changes made (with rule names).
===TEXT===
[PASTE]
===END===

28. The Email Rewriter

Make sent emails sound less robotic.

Rewrite this email to be [WARMER / FIRMER / SHORTER / MORE DIRECT]. Keep all factual content. Subject line stays the same unless it's clearly bad. Max 1 ask per email. Output rewritten email only.
===EMAIL===
[PASTE]
===END===

29. The Long-Doc Summarizer

Stratified summary for long inputs.

Summarize the document below in three layers: (1) ONE-SENTENCE summary, (2) THREE-BULLET summary, (3) FIVE-PARAGRAPH summary covering setup, argument, evidence, counterpoint, conclusion. Each layer self-contained.
===DOC===
[PASTE]
===END===

30. The Quote Puller

Extract pull-quote candidates.

From the text below, extract the 5 most quotable sentences. Criteria: standalone meaning, vivid language, strong opinion, useful for social sharing. Output as a numbered list. For each, add a one-line "why this one" rationale.
===TEXT===
[PASTE]
===END===

4. Marketing & SEO (Prompts 31–40)

31. The Blog Post Outline

Search-intent-aware outlines.

Target keyword: "[KEYWORD]". Search intent: [INFORMATIONAL/COMMERCIAL/TRANSACTIONAL]. Build a blog post outline that beats the top 3 results. Output: title (≤60 chars), meta description (≤155 chars), 8–12 H2 sections, and 3 FAQ questions to target People-Also-Ask.

32. The Meta Tag Generator

Title + description + OG.

Page topic: "[PAGE TOPIC]". Primary keyword: [KEYWORD]. Generate: (1) 3 title tag variants (≤60 chars each, keyword in first 30), (2) 3 meta descriptions (≤155 chars, action verb, includes keyword), (3) one OG description (≤90 chars, hook-focused). Output as JSON.

33. The Twitter Thread Writer

Hook-driven, scannable threads.

Topic: [TOPIC]. Audience: [WHO]. Goal: [TAKEAWAY]. Write a 7-tweet thread. Tweet 1: hook (curiosity or contrarian, ≤220 chars, no emojis). Tweets 2–6: one idea each, ≤270 chars, can use line breaks. Tweet 7: CTA. Number each tweet.

34. The LinkedIn Post Writer

Personal-voice posts.

Topic: [TOPIC]. Tone: [REFLECTIVE/CONTRARIAN/CELEBRATORY]. Write a LinkedIn post: (1) first line is a 1-sentence hook that breaks pattern, (2) blank line, (3) 3–5 short paragraphs telling a story or making a point, (4) one-line closer with a question. Max 1300 chars. No hashtags.

The hardest 80 words to write.

Write the opening 80 words of a newsletter issue about [TOPIC]. Tone: [CASUAL/PROFESSIONAL]. Must include: a concrete hook (not "happy Monday"), a one-line preview of what's inside, and a reason to keep reading. Cut every word that doesn't earn its place.

36. The Hook Variation Generator

For ads, videos, headlines.

Product: [PRODUCT]. Audience pain point: [PAIN]. Generate 20 opening-line hooks in 4 styles, 5 each: (1) shocking stat, (2) story-in-one-sentence, (3) contrarian claim, (4) direct question. Max 15 words per hook. Output as a numbered list with style in brackets.

37. The Schema Markup Writer

Valid JSON-LD on request.

Generate JSON-LD schema markup for a [ARTICLE/PRODUCT/REVIEW/HOWTO/FAQ] page. Page details below. Output: valid JSON-LD inside a `<script type="application/ld+json">` tag. Use schema.org vocabulary. No extra commentary.
===PAGE DETAILS===
[FILL IN]
===END===

38. The Internal Link Suggester

Anchor text + target URL pairs.

New article below. Existing site URLs and their topics also below. Suggest 5–8 internal links to add to the new article. For each: anchor text (3–6 words, natural), target URL, and the exact sentence in the new article where it should be placed.
===NEW ARTICLE===
[PASTE]
===EXISTING URLS===
[PASTE LIST: url — topic]
===END===

39. The Keyword Cluster Builder

Topic clusters from a seed.

Seed keyword: "[KEYWORD]". Build a topic cluster: (1) one pillar topic, (2) 8–12 subtopics, each a candidate H2 or sibling article. For each subtopic, add: search intent type, estimated difficulty (low/med/high), and the kind of content that would win (guide, listicle, comparison, etc.).

40. The Competitor Page Analyzer

What to beat, what to ignore.

URL/content of a top-ranking competitor page below. Analyze: (1) what's good about it (3 bullets), (2) what's weak about it (3 bullets), (3) 5 specific things our version should do differently to win the rank. Be concrete — no "make it better."
===COMPETITOR===
[PASTE]
===END===

5. Productivity & Workflows (Prompts 41–50)

41. The Meeting Notes Distiller

Raw transcript below. Output: (1) one-paragraph summary, (2) decisions made (bulleted), (3) action items as `OWNER — TASK — DUE`, (4) open questions (bulleted), (5) one risk/blocker if any. Skip pleasantries. If a field is empty, write "—".
===TRANSCRIPT===
[PASTE]
===END===

42. The Calendar Planner

I have these tasks: [LIST WITH ROUGH DURATIONS AND DEADLINES]. My deep-work hours are [TIME RANGES]. Generate a tomorrow's calendar: 25-min Pomodoros with 5-min breaks, hardest task first, no more than 4 hours of deep work, buffer for meetings: [LIST MEETINGS]. Output as a timestamped schedule.

43. The Daily Standup Writer

From the activity log below, write a 3-line standup update: yesterday (what shipped), today (what I'm doing), blockers (or "none"). Each line ≤140 chars. No filler. No emojis.
===LOG===
[PASTE COMMITS / TASKS / NOTES]
===END===

44. The Email Triager

Below are the subject lines and senders of [N] unread emails. Sort them into 4 buckets: REPLY-TODAY, REPLY-THIS-WEEK, READ-ONLY, ARCHIVE. For each REPLY-TODAY item, draft a 1–2 sentence reply.
===INBOX===
[PASTE]
===END===

45. The Task Decomposer

Goal: "[GOAL]". Constraint: must be done by [DATE], by [WHO]. Decompose into 3–7 atomic tasks, each ≤2 hours of work, with clear "done when…" criteria. Output as a numbered list. Flag dependencies between tasks.

46. The Decision Matrix

I'm choosing between: [OPTION A], [OPTION B], [OPTION C]. Criteria that matter to me: [LIST CRITERIA WITH WEIGHTS 1–5]. Build a weighted scoring table. Score each option 1–5 per criterion. Output: markdown table, weighted total, one-line recommendation.

47. The Status Report Generator

This week's activity below. Write a stakeholder-facing weekly status report: (1) headline (1 sentence), (2) shipped this week (bullets), (3) in flight (bullets), (4) risks / asks (bullets, or "none"), (5) next week's focus (bullets). Total ≤200 words.
===ACTIVITY===
[PASTE]
===END===

48. The OKR Drafter

Theme: [QUARTERLY THEME]. Team: [TEAM]. Draft 1 objective + 3 key results. Objective must be qualitative and inspirational. KRs must be measurable, time-bound, and have a baseline + target. No vanity metrics. Output formatted as `O: ...` and `KR1/2/3: from X to Y by [DATE]`.

49. The Weekly Reviewer

Below are this week's wins, misses, and blockers. Run a structured weekly review: (1) what worked (2–3 bullets), (2) what didn't (2–3 bullets), (3) one lesson, (4) one thing to do differently next week, (5) one thing to stop doing entirely.
===INPUT===
[PASTE]
===END===

50. The Inbox Zero Coach

I have [N] unread emails. Help me get to zero in 30 minutes. Output a 5-step plan: (1) which categories to bulk-archive first, (2) which folder rules to set up, (3) a script for the 3 email templates I'll need most, (4) a 60-second rule for the rest, (5) what to do tomorrow to keep it at zero.

6. Customer Support & Sales (Prompts 51–60)

51. The Empathetic Support Reply

Customer message below. Reply: (1) acknowledge the specific frustration in their own words, (2) explain what happened in plain language, (3) state what you're doing about it with a timeline, (4) end with a single concrete next step. No "we apologize for any inconvenience." Max 150 words.
===MESSAGE===
[PASTE]
===END===

52. The Firm Support Reply

Customer is asking for [REQUEST] that violates [POLICY]. Reply: (1) one-line empathetic acknowledgment, (2) clear no with the specific policy reason, (3) one alternative we CAN offer, (4) close politely. Do not apologize for the policy. Max 120 words.
===MESSAGE===
[PASTE]
===END===

53. The Refund Decision

Refund request below. Decide: APPROVE, PARTIAL, or DENY. Reasoning in one paragraph, citing: customer history, request validity, business impact. Then draft the customer-facing reply matching the decision. Output: decision, reasoning, reply.
===REQUEST===
[PASTE]
===CUSTOMER HISTORY===
[PASTE]
===END===

54. The Knowledge Base Article

Question we get often: "[QUESTION]". Write a KB article: (1) H1 phrased as the user's question, (2) one-paragraph answer at the top, (3) step-by-step instructions if applicable, (4) "When this doesn't work" troubleshooting section, (5) "Related" section with 2–3 link suggestions. Max 400 words.

55. The Sales Follow-Up

Last contact with prospect: [SUMMARIZE WHAT WAS DISCUSSED, WHEN]. Their stated objection or hesitation: [OBJECTION]. Write a follow-up email that: (1) references something specific from last conversation, (2) addresses the objection with one concrete proof point, (3) proposes a low-friction next step. Max 110 words.

56. The Cold Outreach

Prospect: [NAME, TITLE, COMPANY]. Their company's likely pain point: [PAIN]. Our offer: [OFFER]. Write a cold email: (1) personalized opener referencing something real about them or their company, (2) one-sentence relevance bridge, (3) one specific value claim with a number, (4) low-friction ask (15-min call OR just a reply). Max 90 words. No "I hope this finds you well."

57. The Discovery Question Generator

I'm about to do a discovery call with [TITLE] at a [COMPANY TYPE]. Goal: understand if our [PRODUCT] fits. Generate 12 questions in 4 buckets, 3 each: (1) current state, (2) goals, (3) gaps/pain, (4) decision process. Order them so the call flows naturally. Each question ≤20 words.

58. The Objection Handler

Customer just said: "[OBJECTION]". I'm selling [PRODUCT/SERVICE]. Generate 3 ways to respond: (1) acknowledge + reframe, (2) acknowledge + counter with proof, (3) acknowledge + question back to surface the real concern. Each response ≤60 words. Don't be defensive.

59. The Demo Script

Product: [PRODUCT]. Audience: [TITLE]. Their use case: [USE CASE]. Their top concern: [CONCERN]. Build a 15-minute demo script: (1) 1-min opening (mirror back their problem), (2) 10 min of demo split into 3 chapters with what to show and what to say, (3) 2-min question prompt, (4) 2-min next-steps close.

60. The Renewal Email

Customer renewal in [N] days. Their usage this year: [USAGE METRICS]. Their original goal when signing: [GOAL]. Write a renewal email: (1) lead with a specific outcome they got, (2) name one feature they've underused that could compound the value, (3) renewal CTA. No discounts unless I say so. Max 130 words.

7. Research & Summarization (Prompts 61–70)

61. The Paper TL;DR

Academic paper below. Output: (1) one-sentence TL;DR, (2) the problem in plain language, (3) the method in 2 sentences, (4) the result with the headline number, (5) one caveat, (6) one practical implication. Total ≤180 words.
===PAPER===
[PASTE ABSTRACT + INTRO + CONCLUSION]
===END===

62. The Technical Paper Summary

Paper below. Summarize for a [DOMAIN] practitioner. Cover: prior work it builds on, what's actually new, experimental setup, key results with numbers, threats to validity, where it would and wouldn't generalize. 400–500 words. Markdown headers.
===PAPER===
[PASTE]
===END===

63. The Multi-Doc Synthesizer

Below are [N] sources on [TOPIC]. Produce: (1) the points of consensus across sources (bulleted), (2) the points of disagreement (with which source takes which side), (3) the strongest single claim from each source, (4) one open question none of them answer.
===SOURCES===
[PASTE WITH `SOURCE 1: ...` HEADERS]
===END===

64. The Quote Extractor

From the source below, extract every direct quote attributable to [PERSON/ROLE]. For each: the exact quote (verbatim), context (1 sentence), and page/timestamp if available. Output as a markdown table.
===SOURCE===
[PASTE]
===END===

65. The Counter-Argument Finder

Argument: "[CLAIM]". Steelman the 3 strongest counter-arguments. For each: (1) the counter-claim in one sentence, (2) the strongest evidence supporting it, (3) the kind of person who would find it most persuasive. Be fair to the counter side.

66. The Source Credibility Check

Source below. Rate credibility on 5 dimensions, each 1–5: author expertise, evidence quality, transparency about methods, recency, bias indicators. Provide a one-line justification per dimension. End with an overall verdict: HIGH / MEDIUM / LOW / DO-NOT-CITE.
===SOURCE===
[PASTE URL + KEY CONTENT]
===END===

67. The Glossary Builder

Domain: [DOMAIN]. Document below. Extract every domain-specific term used. For each: term, plain-English definition (≤25 words), and the sentence from the doc where it's used. Sort alphabetically. Output as a markdown table.
===DOC===
[PASTE]
===END===

68. The Timeline Builder

Topic: [TOPIC]. Sources below. Build a chronological timeline of key events. For each event: date (as precise as possible), 1-sentence what happened, source citation. Mark uncertain dates with [~]. Output: markdown table sorted oldest-first.
===SOURCES===
[PASTE]
===END===

69. The Comparison Table

Compare [ITEMS]: [LIST]. Dimensions to compare: [LIST]. For each cell, give a 5–12 word answer based ONLY on the sources below. If a source doesn't cover something, write "—" rather than inventing. Output: markdown table. After the table, name a winner per dimension.
===SOURCES===
[PASTE]
===END===

70. The Briefing Memo

Audience: [DECISION MAKER ROLE]. They have 5 minutes. Topic: [TOPIC]. Sources below. Write a briefing memo: (1) BLUF (Bottom Line Up Front, 2 sentences), (2) background (3 bullets), (3) options (2–4, each with pros/cons), (4) recommendation with rationale, (5) what we need to decide today. Max 400 words.
===SOURCES===
[PASTE]
===END===

8. Learning & Teaching (Prompts 71–80)

71. The ELI5 Explainer

Explain [CONCEPT] to a curious 12-year-old who is good at math but has no domain knowledge. Use one extended analogy that runs through the whole explanation. End with one "and that's why it matters" sentence. Max 200 words. No jargon.

72. The Socratic Tutor

I'm learning [TOPIC] and I think I understand [SPECIFIC POINT]. Don't tell me whether I'm right. Instead, ask me 3 questions that will reveal where my understanding breaks down if it does. Wait for my answers before continuing.

73. The Quiz Generator

Source material below. Generate 10 multiple-choice questions covering its key ideas. 4 options each, 1 correct. Mix difficulty: 4 recall, 4 application, 2 synthesis. Output as JSON: `[{question, options:[], correctIndex, explanation}, ...]`.
===SOURCE===
[PASTE]
===END===

74. The Flashcard Generator

From the source below, generate flashcards in Anki cloze-deletion format. 15 cards. Each card: one fact, one cloze, ≤25 words. Test the most testable, not the most interesting. Output one card per line as `Card N: text with {{c1::cloze}} deleted`.
===SOURCE===
[PASTE]
===END===

75. The Lesson Plan

Topic: [TOPIC]. Audience: [LEARNER LEVEL]. Time: [N] minutes. Build a lesson plan: (1) learning objective (1 sentence), (2) prerequisite knowledge check (3 questions), (3) explanation phase with one example, (4) practice phase with 2 exercises, (5) check-for-understanding (1 question). Include rough timing.

76. The Curriculum Gap Finder

Below is what I've already learned about [TOPIC]. Identify the 5 most important gaps in my knowledge — concepts I'm missing that would block deeper understanding. For each: what it is (1 sentence), why it matters, and the single best resource to learn it (book, paper, tutorial).
===WHAT I KNOW===
[PASTE]
===END===

77. The Code-Along Walkthrough

Teach me [CONCEPT] in [LANGUAGE] by walking me through building a tiny working example. Format: alternating between a code snippet (≤10 lines) and a 2-sentence explanation of what just happened and why. Build incrementally. End with the full working file.

78. The Mental Model Explainer

I want a mental model for [TOPIC]. Give me one — exactly one — model that captures the most important dynamic. Output: (1) name the model, (2) describe it in 100 words, (3) one concrete example of it in action, (4) one place the model breaks down. No "there are many ways to think about this."

79. The Worked Example Generator

Concept: [CONCEPT]. Generate a worked example: (1) state the problem precisely, (2) solve it step by step, with the reasoning for each step in italics, (3) state the final answer, (4) add one "common mistake to avoid" callout. Domain: [MATH/CODE/REASONING/ETC].

80. The Reading List Builder

Goal: become competent in [TOPIC] in [TIMEFRAME]. Current level: [LEVEL]. Build a sequenced reading list of 8–12 items. For each: title, author, format (book/paper/post), estimated time, why it's on the list at that position. Order matters — earlier items should unlock later ones.

9. Creative & Brainstorming (Prompts 81–90)

81. The Naming Generator

I'm naming a [PRODUCT/FEATURE/COMPANY/CHARACTER]. It does [DESCRIPTION]. The vibe should be [VIBE]. Generate 20 names in 4 styles, 5 each: (1) made-up words, (2) compound nouns, (3) metaphors, (4) verbs. For each: the name and a one-sentence rationale. Avoid AI cliches (Nexus, Synergy, Catalyst, etc.).

82. The Tagline Generator

Product: [PRODUCT]. Core promise: [PROMISE]. Audience: [WHO]. Generate 15 tagline options across these registers: 5 declarative, 5 imperative, 5 questioning. Max 8 words each. No "imagine if…" or "the future of…" openers.

83. The Story Premise Generator

Genre: [GENRE]. Constraint: [CONSTRAINT — e.g., must take place in one room, must have an unreliable narrator]. Generate 5 story premises. Each: 1 logline (≤25 words), the central conflict in one sentence, the moment when things change.

84. The Title Page Test

Below is the opening 500 words of a piece of writing. If a reader saw only this, would they keep reading? Rate: hook strength (1–10), clarity of stakes (1–10), voice distinctiveness (1–10). Then identify the single weakest sentence and propose a stronger replacement.
===OPENING===
[PASTE]
===END===

85. The SCAMPER Ideator

Subject: [PRODUCT/PROCESS/CONCEPT]. Apply SCAMPER. For each letter (Substitute, Combine, Adapt, Modify, Put to other uses, Eliminate, Reverse), generate 2 ideas. Each idea: one sentence describing the change, one sentence on why it could matter. Skip ideas that are obvious or trivial.

86. The Lateral Thinker

Problem: [PROBLEM]. Most obvious solutions: [LIST 2–3 IF YOU HAVE THEM]. Generate 5 lateral solutions — ideas that come from a completely different domain or first principle. Each: the idea (1 sentence), the domain it borrows from, what would have to be true for it to work.

87. The Crazy 8s Prompt

Design challenge: [CHALLENGE]. Constraint: [ONE HARD CONSTRAINT]. Generate 8 wildly different concept directions in 1 round. Each: name, one-sentence description, one-line rationale. Range from safest to weirdest. Don't self-censor — bad ideas welcome.

88. The Pitch Deck Outliner

Company: [STARTUP]. Audience: [SEED / SERIES A / GROWTH INVESTORS]. Build a 10-slide pitch deck outline. For each slide: title, one-line of what to argue, and the single most important number or visual to include. Story arc must build to the ask.

89. The World-Building Prompt

Setting: [TIME, PLACE, GENRE]. Generate a world bible: (1) one defining physical fact about this world, (2) one social rule everyone follows, (3) one technology or magic system rule, (4) the central tension, (5) the kind of conflict that drives stories here. Each in 2–3 sentences.

90. The Dialogue Writer

Scene: [SCENE]. Characters: [A] who wants [GOAL A], [B] who wants [GOAL B]. They cannot both win. Write the scene as dialogue only, no action lines. Subtext > text. End on a line that shifts the power balance.

10. Multimodal: Vision & Audio (Prompts 91–100)

These leverage Flash's strong multimodal input — images, screenshots, charts, audio. Paste the prompt and attach the file.

91. The Image Describer

Describe the attached image in three layers: (1) one-sentence summary, (2) factual contents (objects, people, setting, text visible), (3) interpretive layer (mood, likely context, audience). Total ≤200 words. Flag anything ambiguous with [uncertain: ...].

92. The Diagram-to-Code

The attached image is a [DIAGRAM TYPE: flowchart / ERD / sequence diagram / state machine]. Convert it to [MERMAID / PLANTUML / DOT] source. Preserve every node, edge, label, and direction. Output code block only.

93. The Whiteboard-to-Doc

The attached image is a photo of a whiteboard from a meeting on [TOPIC]. Convert to a clean markdown document: extract every bullet, arrow, and diagram into prose + structured lists. Resolve handwriting ambiguities with [?]. Note which sections seemed most emphasized (boxed, underlined, starred).

94. The Receipt-to-JSON

Attached image is a receipt. Extract: vendor name, date, line items (name, qty, unit price, total), subtotal, tax, total, payment method, last 4 of card. Output as JSON. If a field is illegible, use null and add a `_warnings` array listing which fields were uncertain.

95. The Chart-to-Summary

The attached image is a chart. Output: (1) chart type, (2) what it's measuring (axes, units), (3) the headline finding in one sentence, (4) the most surprising data point, (5) one question the chart raises but doesn't answer. Do not invent precise numbers — describe shapes and rough magnitudes only.

96. The Screenshot-to-Bug-Report

Attached screenshot shows a UI issue I'm reporting. Write a bug report: (1) summary (1 sentence), (2) steps to reproduce (inferred + ask me for what's missing), (3) expected behavior, (4) actual behavior visible in screenshot, (5) severity guess with rationale. Output as markdown.

97. The Slide Deck Extractor

Attached image is a slide from a presentation. Extract: title, subtitle if any, every bullet/text element verbatim, presenter notes if visible, and a 1-sentence guess at the slide's purpose in a deck. Preserve the visual hierarchy in the output.

98. The Audio Transcript Cleaner

Raw transcript below (auto-generated, messy). Clean it: remove filler words (um, like, you know), fix obvious mis-hearings using context, add paragraph breaks at topic shifts, add speaker labels if inferable. Do NOT paraphrase — keep exact wording where intelligible. Output cleaned transcript only.
===TRANSCRIPT===
[PASTE]
===END===

99. The Audio-to-Action-Items

Attached audio is a [MEETING / VOICE NOTE / CALL]. Extract action items only. Format: `OWNER — TASK — DUE`. If owner or due date is implicit, infer with [inferred]. Skip discussion that didn't lead to a decision. Order by mentioned-priority. Max 15 items.

100. The Video Frame Critique

Attached frame is from a [SHORT-FORM VIDEO / AD / TUTORIAL] at the [HOOK / MID / OUTRO] position. Critique it for that role: (1) what's working, (2) what's weak, (3) one specific change that would improve performance. Be concrete — no "make it more engaging."

11. Bonus: Agentic & Tool-Use (Prompts 101–110)

These are the prompts we use when Flash is one model in a larger agent loop — orchestrated by something like a planner-model on Claude Opus or a workflow tool like Genspark.

101. The Plan-Then-Execute Agent

You are an agent. Goal: [GOAL]. Available tools: [LIST]. Output a numbered plan first, max 7 steps. Wait for me to type "GO" before executing. When executing, do one step per response, output: which step, what tool you'll call, why, and what you expect back. After the result, decide: continue / revise plan / done.

102. The Tool Router

You are a routing layer. Given a user message and the tool registry below, output ONLY a JSON object: `{"tool": "tool_name", "args": {...}, "confidence": 0.0-1.0, "reasoning": "one sentence"}`. If no tool fits, return `{"tool": null, "reasoning": "..."}`. Never call the tool yourself.
===TOOLS===
[PASTE TOOL SCHEMAS]
===USER MESSAGE===
[PASTE]
===END===

103. The Self-Critique Loop

Below is a draft answer to "[ORIGINAL QUESTION]". Critique it as a skeptical expert: (1) 3 specific weaknesses, (2) 1 factual claim that should be verified, (3) 1 missing perspective. Then rewrite the answer addressing every critique. Output: critique first, then rewrite.
===DRAFT===
[PASTE]
===END===

104. The JSON-Only Enforcer

You are a JSON API. The user will describe what they want. You will reply with valid JSON only — no markdown, no code fences, no preamble, no explanation. If the request is impossible or ambiguous, return `{"error": "...reason..."}`. Schema for valid responses: [SCHEMA]. User request follows.

105. The Function Call Schema Builder

Function name: [NAME]. What it does: [DESCRIPTION]. Generate a function-calling schema in [OPENAI / ANTHROPIC / GEMINI] format. Include: every required parameter with type and description, every optional parameter with default, return type description, and one example call.

106. The Multi-Agent Orchestrator

Task: [TASK]. Available agents: [LIST: name → specialty]. Decompose the task and assign sub-tasks. Output as JSON: `{"workflow":[{"agent":"X","task":"...","depends_on":[]}], "estimated_steps": N}`. Make dependencies explicit. Do not execute — just plan.

107. The RAG Prompt

You are answering using ONLY the context below. Rules: (1) if the answer isn't in the context, say "I don't know based on the provided sources" — do not use outside knowledge, (2) cite the source ID for every claim like [src:3], (3) if sources disagree, say so and present both sides.
===QUESTION===
[USER QUESTION]
===CONTEXT===
[RETRIEVED CHUNKS WITH IDs]
===END===

108. The Long-Context Loader

I'm about to paste a [DOCUMENT TYPE] of ~[N] tokens. First, acknowledge the upload with the word "LOADED" and nothing else. Then I'll ask questions about it. When answering, always cite the section header or page number you're drawing from. Treat sections not in the doc as "not addressed."

109. The Cost-Aware Router

Task description: "[TASK]". Available models: Flash (cheap, fast, weaker reasoning), Pro (mid, balanced), Opus (expensive, strongest reasoning). Recommend the cheapest model that's likely to succeed. Output: model name, confidence (0–1), reasoning (1 sentence), and what to escalate to if the cheaper model fails.

110. The Confidence Rater

Below is my draft answer to "[ORIGINAL QUESTION]". Rate your confidence in each factual claim from 0–100. Output as a markdown table: claim | confidence | what would change your confidence. Then give the overall confidence in the answer as a whole and identify the single claim most worth verifying first.

How to Chain These Into Pipelines

Single prompts get you single answers. Chains get you systems. Five patterns we use weekly:

Plan → Execute → Critique: Use Prompt 5 (Refactor Planner) → Prompts 1–4 (Coding) → Prompt 103 (Self-Critique). Three model calls, dramatically better output than any single call.
Research → Synthesize → Brief: Prompt 63 (Multi-Doc Synthesizer) → Prompt 65 (Counter-Argument) → Prompt 70 (Briefing Memo). The decision memos we send executives all run through this exact chain.
Extract → Validate → Load: Prompt 94 (Receipt-to-JSON) → Prompt 14 (Data Validator) → your ETL pipeline. Flash is excellent as the extraction layer in a larger data system.
Cold outreach scale-up: Prompt 56 (Cold Outreach) running per-prospect → Prompt 27 (Style Enforcer) for brand-voice consistency → Prompt 110 (Confidence Rater) to gate which ones get sent vs which need a human eye.
Long-doc Q&A: Prompt 108 (Long-Context Loader) → Prompt 107 (RAG Prompt) with the doc as context → Prompt 110 (Confidence Rater) on every answer. This is most enterprise "chat with your docs" features in 100 tokens.

How to Use This Cheat Sheet

Why Gemini 3.5 Flash Specifically

10 Meta-Rules That Make Every Flash Prompt Better

1. Coding & Engineering (Prompts 1–10)

1. The Function Generator

2. The Bug Finder

3. The Code Reviewer

4. The Test Generator

5. The Refactor Planner

6. The Regex Builder

7. The Docstring Writer

8. The API Client Generator

9. The SQL Writer

10. The Commit Message Writer

2. Data & Analytics (Prompts 11–20)

11. The CSV Summarizer

12. The SQL Query Optimizer

13. The JSON-to-Schema Converter

14. The Data Validator

15. The Anomaly Detector

16. The Pivot Summary

17. The Plotting Prompt

18. The Cohort Analyzer

19. The Funnel Analyzer

20. The Data Cleaner

3. Writing & Editing (Prompts 21–30)

21. The Tone Matcher

22. The Length Editor

23. The Clarity Rewriter

24. The Translator

25. The Headline Generator

26. The Outline Builder

27. The Style Enforcer

28. The Email Rewriter

29. The Long-Doc Summarizer

30. The Quote Puller

4. Marketing & SEO (Prompts 31–40)

31. The Blog Post Outline

32. The Meta Tag Generator

33. The Twitter Thread Writer

34. The LinkedIn Post Writer

35. The Newsletter Intro

36. The Hook Variation Generator

37. The Schema Markup Writer

38. The Internal Link Suggester

39. The Keyword Cluster Builder

40. The Competitor Page Analyzer

5. Productivity & Workflows (Prompts 41–50)

41. The Meeting Notes Distiller

42. The Calendar Planner

43. The Daily Standup Writer

44. The Email Triager

45. The Task Decomposer

46. The Decision Matrix

47. The Status Report Generator

48. The OKR Drafter

49. The Weekly Reviewer

50. The Inbox Zero Coach

6. Customer Support & Sales (Prompts 51–60)

51. The Empathetic Support Reply

52. The Firm Support Reply

53. The Refund Decision

54. The Knowledge Base Article

55. The Sales Follow-Up

56. The Cold Outreach

57. The Discovery Question Generator

58. The Objection Handler

59. The Demo Script

60. The Renewal Email

7. Research & Summarization (Prompts 61–70)

61. The Paper TL;DR

62. The Technical Paper Summary

63. The Multi-Doc Synthesizer

64. The Quote Extractor

65. The Counter-Argument Finder

66. The Source Credibility Check

67. The Glossary Builder

68. The Timeline Builder

69. The Comparison Table

70. The Briefing Memo