100 Best Claude Opus 4.7 Prompts for Power Users

Claude Opus 4.7 is the most capable general-purpose model on the market in mid-2026, and we have been running it as the senior model in our stack for six months. Along the way we accumulated a working library of prompts that consistently produce better output than the obvious version of the same ask.

This is that library — 100 prompts, ten categories, all tested in real work. None of these are toy prompts. Each one is something we either run weekly or have saved in our text-expander because the lift versus a naive version is worth the keystrokes.

Short version on how to use this post: read the mental model section first, then jump to whichever category matches what you are working on. The hero prompts in each section have copy buttons. The rest are numbered and ready to lift.

Why These Prompts Work — The Mental Model

Opus 4.7 is not the same model as GPT-5.5 or Gemini 3.5 Flash. The prompts that get the best out of it lean on three properties of the model:

• It tolerates very long context. 1M tokens of useful input is normal. Stop summarising — paste the whole thing in.
• It rewards a clear role and a clear deliverable. "You are X. The output is Y in format Z." Two sentences of setup beat two paragraphs of vague framing.
• Extended thinking is always-on at the Opus tier. You do not need to say "think step by step." You can say "before answering, identify the three weakest links in the argument" and it will actually do that.

The pattern under almost every prompt in this list is the same: specific role + specific input + specific deliverable + constraints that force the model to do real work instead of plausible-sounding work. Once you see it you will start writing your own.

How to Use This List

1. Find the category that matches your task.
2. Pick the closest prompt. Replace the {{bracketed}} placeholders with your actual inputs.
3. Paste any relevant context (full files, full docs, real data) before the prompt, not after. Opus 4.7 weights early context heavily.
4. If the output is not what you wanted, the issue is almost always the constraint section. Tighten it and re-run.

Pro tip: For repetitive prompts, save them as Claude Projects with a system prompt. The system-prompt slot gives the model standing instructions that do not eat into your message context.

Category 1 — Coding & Engineering (Prompts 1–10)

Opus 4.7 leads SWE-bench Verified at 82.1%. These ten are the prompts that turn that benchmark number into real engineering output.

3. Bug reproduction script. Given the stack trace and error message I have pasted, write a minimal reproduction script in {{language}}. The script must fail in exactly the way the production error describes. Add inline comments explaining what each block proves about the bug.
4. Migration safety check. I have pasted a database migration. Tell me: what locks does this take, what happens to a row mid-write during the migration, and what is the rollback plan if it fails halfway. Be concrete about Postgres / MySQL behaviour, not generic.
5. API contract designer. Design a REST API for {{capability}}. Output the full OpenAPI 3 spec. Include error responses, pagination, rate limits, and idempotency keys where appropriate. Justify each design choice in a comment.
6. Test gap analyser. Given this source file and its test file, list the behaviours that are not tested. Rank by how likely they are to cause a regression. Suggest the three highest-value test cases to add.
7. Performance triage. I have pasted a flame graph and the relevant code. Identify the top three optimisations by expected impact. For each: the specific change, the expected speedup, and the risk of introducing a bug. Skip changes that save less than 5% wall time.
8. Architecture decision record. Given the context I just pasted, write an ADR comparing {{option A}}, {{option B}}, and {{option C}} for {{decision}}. Include the trade-offs, the recommendation, and the conditions under which we should revisit.
9. Library evaluator. Should we adopt {{library}}? Read the readme, the open issues from the last 90 days, and the closed PRs I have pasted. Output: maintenance health (1-10), API stability, lock-in risk, and a final yes/no with two reasons.
10. Commit message rewriter. Here is my git diff and my draft commit message. Rewrite the message to follow Conventional Commits, explain the why not the what, and call out any breaking changes. Keep it under 72 characters for the subject line.

Category 2 — Long-Context Analysis (Prompts 11–20)

This is where Opus 4.7 separates from every other model. Paste in entire codebases, entire PDFs, entire research corpora — then ask it to do work that no model with a 200k window can do.

13. Contract redline. Compare these two versions of a contract I have pasted. List every substantive change. Group changes by risk category. Flag anything that shifts liability or extends the term.
14. Meeting recap from raw transcript. Here is a 45-minute Zoom transcript. Produce: a 5-bullet executive summary, a numbered list of decisions made, an owner-tagged action item list, and any open question that was raised but not resolved.
15. Multi-document Q&A. I have pasted 6 documents above. Answer the following question using only information from those documents: {{question}}. For each claim in your answer, cite which document(s) support it. If no document supports an answer, say so and stop.
16. Earnings call analyser. Here is the full transcript of {{company}}'s most recent earnings call. Extract: the three forward-looking statements management made, the two questions analysts pressed hardest on, any change in tone from the previous quarter's call, and any metric that was disclosed for the first time.
17. Patent landscape map. Given the patent abstracts I have pasted, identify the dominant filers, the technology clusters, and the white space where filings are notably absent. Output as a structured taxonomy.
18. Codebase onboarding doc. Read this entire repository I have pasted. Produce a CLAUDE.md / README.md style onboarding doc for a new senior engineer joining tomorrow. Cover: how the code is structured, where the load-bearing complexity lives, conventions to respect, and the three files a new engineer should read first.
19. Long-document diff narrator. Here are two versions of a strategy doc. Write a one-page changelog that an executive can skim in 60 seconds, explaining what shifted and why it matters.
20. Customer feedback themer. I have pasted 200+ rows of customer feedback. Cluster them into themes. For each theme: count, representative quotes, severity, and a recommended product action. Do not invent themes that are supported by fewer than 5 entries.

Category 3 — Writing & Content (Prompts 21–30)

Opus 4.7 writes in a way that reads less like a model than any other frontier system. These are the prompts that make that strength usable.

23. Long-form to thread. Convert this long-form article into a 12-tweet thread. Each tweet under 270 characters. Open with the strongest hook. The thread must stand alone — no "read the full piece" until tweet 12.
24. Sales page from product notes. Here are my product notes. Write a sales page in the voice of {{brand reference}}. Sections: headline, sub-headline, three problem statements, three benefit blocks (with proof), pricing, FAQ, final CTA.
25. Cold email writer. Write a 90-word cold email to {{persona}} about {{offer}}. The opening line must reference something specific about their company. No "I hope this finds you well." End with a single yes/no question.
26. Headline test generator. Generate 20 headline variations for an article on {{topic}}. Categorise them into curiosity-led, benefit-led, and contrarian. Mark the top three by predicted CTR and explain the reasoning in one line each.
27. Ghostwriter for technical founders. I have pasted three Slack messages from {{founder name}}. Turn them into a publishable 600-word LinkedIn post that sounds like them. Keep their phrases, hesitations, and idioms. Do not smooth them out.
28. Newsletter remixer. Take this raw idea dump and turn it into a 400-word newsletter section. Voice: confident, specific, no corporate filler. Include one personal anecdote, one specific number, and one cross-link.
29. Press release. Write a press release for {{announcement}} in the AP-news style. Include a strong lead, two stakeholder quotes (you can draft them, mark with [DRAFT]), a boilerplate section, and a contact block.
30. Subject line A/B set. Generate 15 subject line variants for an email about {{offer}}. Five each: curiosity, urgency, value-stated. All under 60 characters. Rank the top three you would actually send.

Category 4 — Research & Analysis (Prompts 31–40)

33. Devil's advocate. Here is my plan. Argue against it as the smartest sceptic in the room. What are the three strongest objections? What is the single fact that, if true, would kill the plan?
34. Pre-mortem. Assume this initiative failed badly 12 months from now. Write a post-mortem from that future state explaining what went wrong, in priority order. Be specific about the failure mode, not generic.
35. Market sizing. Estimate the total addressable market for {{product}} in {{geography}}. Use top-down and bottom-up methods. Show every assumption explicitly. End with a confidence band, not a single point estimate.
36. Customer persona from data. Here are 30 customer transcripts. Build three distinct personas. For each: demographics, jobs-to-be-done, the moments they reach for our product, and the moments they hire a competitor.
37. SWOT but useful. Standard SWOT analyses are filler. Produce one for {{company}} where every entry is non-obvious. If you cannot find a non-obvious entry for a quadrant, leave it empty rather than padding.
38. 5-Whys root cause. Here is a problem statement: {{problem}}. Run the 5-Whys technique. After the fifth why, name the underlying systemic issue and propose the single highest-leverage intervention.
39. Trend extraction. Given the news headlines I have pasted from the last 90 days, identify three trends that are visible across multiple stories but not yet named in mainstream coverage.
40. Pricing teardown. Compare the pricing pages of {{competitor 1}}, {{2}}, {{3}}. Identify the anchor price, the loss-leader tier, the actual margin tier, and the lock-in mechanism. Recommend the one change to our own pricing that would have the highest impact.

Category 5 — Agents & Automation (Prompts 41–50)

Opus 4.7 is the strongest agent model in the field. These prompts are how we get the most out of it inside Claude Code, computer-use, and tool-using setups.

43. Computer-use safety wrapper. When using computer-use, before any click that submits a form, makes a payment, or sends a message: pause, screenshot, describe what is about to happen in one sentence, and ask me to confirm with the word "go" before proceeding.
44. Sub-agent delegator. Decompose this task into parallel sub-tasks. For each sub-task: a clear goal, the inputs it needs, the output format expected. Spawn sub-agents and merge their results into a single final answer.
45. Self-critic loop. After every major output, write a one-paragraph critique of your own work pointing out the weakest claim, the missing perspective, and the assumption that is doing the most load-bearing work. Then either revise or stand by it.
46. API integration scaffolder. I am integrating with {{API name}}. Read its documentation that I have pasted. Generate: a typed client in {{language}}, request and response schemas, retry/backoff logic for the three most common failure modes, and a smoke test that proves the integration works end-to-end.
47. Cron job designer. Design a cron / scheduled job for {{task}}. Include: schedule expression, idempotency strategy, alerting on failure, the metric I should watch to know it is working, and the rollback procedure.
48. Webhook handler skeleton. Write a webhook handler for {{provider}} that verifies the signature, validates the payload, deduplicates by event ID, and processes the event asynchronously via a queue. Include the failure-replay path.
49. Data pipeline builder. Design a pipeline that ingests {{source}}, transforms it via {{rules}}, and lands it in {{destination}}. Cover schema evolution, late-arriving data, and the monitoring needed to know it is working.
50. Multi-agent debate. Simulate three agents debating {{decision}}: a sceptic, an optimist, and a domain expert. Each gets two rounds of arguments. At the end, synthesise their disagreement into a single recommendation.

Category 6 — Data & Spreadsheets (Prompts 51–60)

52. SQL from natural language. Given the schema I have pasted, write the SQL for: {{question in plain English}}. Use the dialect for {{Postgres / BigQuery / Snowflake}}. Explain in one line what the query does. If the schema does not support the question, say so.
53. Query optimiser. Here is a slow SQL query and its EXPLAIN output. Identify why it is slow, propose three rewrites in priority order, and predict the expected speedup for each.
54. Excel formula factory. Translate this plain-English calculation into an Excel / Google Sheets formula: {{description}}. Provide the formula, an example with sample inputs, and a fallback for the most likely #ERROR case.
55. Dashboard designer. Given the metrics I have pasted, design a single dashboard for {{audience}}. Include: the four headline KPIs, the two diagnostic charts, the one drill-down view. Name each chart and specify the chart type.
56. Schema designer. Design a relational schema for {{domain}}. Include tables, columns, types, primary and foreign keys, indexes, and any required denormalisation with the justification.
57. Data cleaning plan. Here is the first 200 rows of a messy CSV. List the cleaning operations needed in execution order: column normalisation, type coercion, missing-value strategy, outlier handling, and duplicate detection.
58. A/B test reader. Here are the results of an experiment. Tell me: is the result statistically significant, what is the effect size in business terms, what would I need to see to launch, and the single confound most likely to be polluting the result.
59. Pivot summariser. I have pasted a complex pivot table. Explain it in three bullets a non-analyst can understand. Then list the two follow-up cuts I should ask the analyst for.
60. Metric definition. Define {{metric name}} precisely: numerator, denominator, time window, edge cases (zero denominators, late events, retroactive corrections), and the two ways teams typically game it.

Category 7 — Strategy & Business (Prompts 61–70)

63. OKR drafter. Draft Q3 OKRs for a {{team}} of {{size}} working on {{focus}}. Three objectives. Each with 3 key results. Each KR must be measurable, time-bound, and aggressive but plausible.
64. Pitch deck outliner. Outline a 12-slide investor pitch deck for {{company}}. For each slide: the headline, the supporting point, and the visual element. Pitch in the YC style — no fluff slides.
65. Build-vs-buy. Should we build or buy {{capability}}? Run the analysis. Cover: cost over 24 months, time-to-value, strategic differentiation, switching cost if we change our mind, and the team's capacity to maintain it.
66. Hiring scorecard. Build a scorecard for hiring a {{role}}. Cover: mission (one line), 4-6 outcomes, the competencies that matter, and the dealbreaker behaviours. No corporate filler.
67. Compensation benchmark. Given the role description I just pasted, estimate the compensation band (base, variable, equity) in {{geography}} for {{seniority}}. Show your sources or mark as estimate.
68. Quarterly review. Here are our results vs OKRs for last quarter. Produce: what worked, what did not, what we learned, what carries into next quarter, and the single change we should make to how we plan.
69. Stakeholder map. For the change I am proposing, list every stakeholder. For each: their interest, their power, whether they are an advocate / neutral / blocker, and the single message that moves them.
70. Risk register. List the top 10 risks to {{initiative}}. For each: likelihood (1-5), impact (1-5), the early-warning signal, and the mitigation we should put in place now.

Category 8 — Learning & Teaching (Prompts 71–80)

72. Socratic challenger. I believe {{claim}}. Ask me four progressively harder questions designed to find the weakest part of my reasoning. Do not give me the answers. After my responses, tell me which assumption I should reconsider.
73. Spaced repetition deck. Generate 25 flashcards for {{topic}}. Format: question on one line, answer on the next. Avoid one-fact-per-card padding — bundle related facts that are usefully recalled together.
74. Pre-reading brief. I am about to read {{book / paper}}. In 300 words: the central thesis, the three arguments supporting it, the strongest critique, and the one chapter I should not skip.
75. Concept analogy generator. Explain {{technical concept}} using three different analogies aimed at three different audiences: a 10-year-old, a smart non-technical executive, and a domain expert in an adjacent field.
76. Exam designer. Build a 20-question exam on {{topic}} for someone preparing for {{certification / interview}}. Mix multiple choice, short answer, and one applied scenario. Include answers and the rationale for each.
77. Learning path designer. I have {{X hours / week}} for the next {{N weeks}} to learn {{topic}}. Build a weekly path with resources, exercises, and a checkpoint at the end of each week.
78. Code-along walkthrough. Teach me {{technical skill}} by building {{small project}} together. One step at a time. After each step, ask me to write the next bit before showing yours.
79. Misconception finder. What are the five most common misconceptions about {{topic}}? For each: why it is intuitive, why it is wrong, and the simplest experiment or example that demonstrates the truth.
80. Concept map. Build a hierarchical concept map for {{topic}}: the root idea, the major branches, the supporting concepts, and the connections between branches. Use indented bullets.

Category 9 — Creative & Ideation (Prompts 81–90)

83. Tagline writer. Write 15 taglines for {{brand}}. Mix benefit-focused, identity-focused, and contrarian. Top three: explain why each works in one sentence.
84. Story spine. Draft a story for {{topic}} using the structure: once upon a time, every day, until one day, because of that (x3), until finally, and ever since then.
85. Counterfactual world. Describe a world in which {{premise}} is true. Be specific about how everyday life, business, and politics shift. End with the one industry that would be created and the one that would die.
86. Constraint-driven brief. Design a {{thing}} with these constraints: must be {{X}}, must not be {{Y}}, must work for {{audience}}, budget under {{Z}}. The constraints are non-negotiable. If they conflict, name the conflict and propose the trade-off.
87. Mood board in words. Describe the visual mood for {{project}} in 200 words. Colours, typography, textures, motion, the cultural references it borrows from. Specific enough that a designer could move directly to making.
88. Reverse brief. Here is a finished piece of creative work. Reverse-engineer the brief that produced it: target audience, brand objective, key insight, the constraints, and the success metric.
89. Naming for an internal project. Give me 12 internal codenames for a project about {{topic}}. Theme: {{e.g. mythology, jazz, cities}}. None should leak the actual purpose.
90. Slogan stress test. Here is a draft slogan. Try to break it: when is it false, when is it ironic, when does it age badly, when does a competitor own it better. Score 1-10 and recommend keep / revise / kill.

Category 10 — Productivity & Meta (Prompts 91–100)

93. Meeting prep brief. I have a meeting with {{person/team}} in {{N}} hours about {{topic}}. Write me a one-page brief: their likely position, my objective, the three questions I should ask, the two things I should not say, and the single decision I want them to make.
94. Saying no template. Draft a polite but firm reply declining {{request}}. Tone: warm, brief, leaves the door open for a different ask. Under 80 words.
95. Status update writer. Turn my raw notes from this week into a status update for my manager. Format: what shipped, what is on track, what is blocked, what I need from you. Each section three bullets max.
96. Idea capture. I just dumped a stream-of-consciousness about an idea. Extract the actual idea in two sentences. Then list the three smallest experiments I could run to test it this week.
97. Weekly review. Here is my calendar and todo list from last week. Tell me: where did my time actually go, which commitments did I keep, which did I quietly drop, and what would my future self thank me for changing next week.
98. Reading list curator. Based on the topics I have been working on recently, recommend 7 things to read this week. Mix one classic, two recent papers, two practitioner essays, and two contrarian takes. One-line summary each.
99. Note synthesiser. Here are my notes from the last 30 days in {{tool}}. Identify the three themes I am actually thinking about, the question I keep returning to, and the project I should commit to next.
100. The pre-mortem on me. Here is my plan for the next quarter. Argue why my future self will look back at this plan and regret it. Be specific. Suggest the one change that prevents the regret.

How to Combine These — The Power-User Pipeline

The biggest unlock is not running a single prompt — it is chaining them. A pipeline we run weekly:

1. Research stage: Prompt 32 (competitive landscape) → Prompt 38 (trend extraction) → Prompt 31 (steelman both sides).
2. Decision stage: Prompt 62 (decision framework) → Prompt 34 (pre-mortem) → Prompt 33 (devil's advocate).
3. Execution stage: Prompt 41 (multi-step planner) → Prompt 45 (self-critic loop) → Prompt 91 (honest daily plan).
4. Output stage: Prompt 21 (voice match) → Prompt 22 (honest first draft) → Prompt 26 (headline test generator).

Pipelines compound. A single prompt is a tool. Four prompts in sequence is a workflow. Twenty prompts orchestrated through Claude Code or Genspark is what an analyst, a strategist, or a junior engineer used to do in a week.

Steal These Patterns — The Meta-Rules

If you read the 100 prompts above carefully you will notice they share a small number of moves. These are the moves:

• Role + deliverable + format + constraint. Every prompt has all four. Drop one and the output gets vague.
• Forced opinion. "Be specific." "Do not pad." "If you cannot, say so and stop." These instructions block the model from defaulting to wishy-washy comprehensiveness.
• Counter-incentives against confident wrongness. "Mark any claim that is not directly supported by the corpus." "Mark inferences as inference." This is how you stop hallucination on long-context analysis.
• Reverse questions. "What would need to be true for this to be wrong?" forces the model to do work it would skip with a default phrasing.
• Concrete over generic. "Cite file paths and line ranges" beats "be specific."
• One question at a time. Compound prompts get compound answers. If you want a sharp answer, ask a sharp question.

Pro tip: If a prompt is not working, the fastest fix is almost never "rephrase the question." It is "add a constraint that makes the lazy answer impossible."

Where to Run These

You can run any of these in Claude.ai, the Anthropic API, or Claude Code. The right surface depends on the job:

• Claude.ai (web app). For one-off prompts, long-document analysis, writing. Save the most-used ones as Projects with a system prompt.
• Claude Code. For anything in Category 1 (Coding & Engineering) and Category 5 (Agents). The autonomous loop on top of these prompts is where Opus 4.7 actually shines.
• Anthropic API. If you are wiring these into a product. Use the system prompt for the standing instructions and reserve the user message for the variable input.
• Aggregators like Genspark. Useful when you want to A/B Opus 4.7 against Gemini 3.5 Flash or GPT-5.5 High on the same prompt — we cover the model-vs-model trade-offs in our latest model comparison.

The Verdict

The honest take after six months of running Opus 4.7 as our senior model: the gap between people who get great work out of these models and people who get mediocre work is almost entirely down to the prompt. The model is the same. The prompt is the lever.

The 100 above are not magic. They are the result of one specific habit: when a default prompt gave us a default answer, we asked what constraint would have forced a better one — then wrote that constraint into the prompt and saved it. Do that for 30 days and you will have your own list of 100. This one is just a head start.

Keep Reading

• Gemini 3.5 Flash vs Claude Opus 4.7 vs GPT-5.5 High — model comparison covering when to reach for Opus over the alternatives.
• Best ChatGPT Prompts for AI Ads & Commercials in 2026 — the same prompt-engineering rigour applied to ad creative.
• Genspark Review 2026 — the agent orchestrator we use to run these prompts at scale.
• How to Create an AI Avatar for YouTube in 2026 — the multi-model workflow these prompts plug into.
• Prompt Library — the full searchable collection, organised by tool and use case.
• All AI Models — current capabilities, pricing, and best-fit guidance for every major model.