Plan → Execute → VerifyAgents & reasoning
Turns any chat model into an agent that plans, works in steps and checks its own output.
When to use: Multi-step tasks you want done end-to-end: a plan, a document, a small project.
You are an autonomous task agent. Goal: [DESCRIBE THE GOAL].
Context and materials: [PASTE DATA, LINKS OR NOTES].
Work in this loop:
1. PLAN — break the goal into 3–7 concrete steps. State your assumptions.
2. EXECUTE — complete each step in order. Show the result of each step briefly.
3. VERIFY — after the last step, check the result against the goal. List anything missing, wrong or uncertain and fix it.
4. DELIVER — give the final result in a clean, ready-to-use form.
Rules: do not ask me questions unless a step is truly blocked; make reasonable assumptions and label them. Stop and ask before anything irreversible (sending, deleting, paying).
Tree of ThoughtsAgents & reasoning
Forces the model to generate several solution paths, score them and pursue the best one.
When to use: Problems with several possible approaches where the first idea is rarely the best.
Problem: [DESCRIBE THE PROBLEM].
Think like a team of three independent experts.
1. Each expert proposes a different approach (A, B, C) in 3–4 sentences.
2. Score every approach 1–10 on: effectiveness, cost/effort, risk, speed. Show a table.
3. Each expert critiques the other two approaches in one sentence.
4. Pick the best approach — or combine the strongest parts — and explain why in 3 bullet points.
5. Develop the chosen approach into a concrete step-by-step plan with the first action I can take today.
Reflexion · self-critiqueWriting & communication
The model writes, then critiques its own draft like a harsh editor, then rewrites.
When to use: Important texts: offers, landing pages, applications, emails to decision-makers.
Task: [WHAT SHOULD BE WRITTEN — e.g. proposal, landing page copy, email].
Audience: [WHO READS IT]. Goal: [WHAT THEY SHOULD DO OR FEEL].
Step 1 — Write a first draft.
Step 2 — Switch roles: you are a demanding senior editor. List the 5 biggest weaknesses of the draft (clarity, persuasion, structure, tone, missing facts). Be specific and blunt.
Step 3 — Rewrite the text fixing every weakness.
Step 4 — Give a one-line summary of what changed and why the new version is better.
Output only steps 3 and 4 unless I ask for the full process.
Evidence & confidenceResearch & analysis
Separates facts from assumptions and rates how sure the model is about every claim.
When to use: Questions where you must know how certain the answer is — markets, regulations, technologies.
Act as a senior research analyst. Question: [YOUR RESEARCH QUESTION].
Deliver a brief with:
1. Bottom line — the answer in 3 sentences.
2. Key findings — 5–8 bullets. Tag each: [FACT], [ESTIMATE] or [ASSUMPTION], plus confidence High / Medium / Low.
3. What would change the conclusion — the 3 facts that, if different, flip the answer.
4. Unknowns — what you could not determine and how I could find out.
5. Sources — where each fact comes from. If you are not sure a source exists, say so instead of inventing one.
Never present a guess as a fact.
Pre-mortemBusiness & decisions
Imagines the project already failed and works backwards to find the risks early.
When to use: Before launching a project, product, event or migration.
Project / decision: [DESCRIBE IT, INCLUDING GOAL, TIMELINE AND BUDGET].
It is 12 months from now and this project has clearly failed.
1. Write the 7 most likely reasons it failed — be concrete, not generic.
2. For each: probability (1–5), impact (1–5), and an early warning sign I could watch for.
3. Sort by probability × impact.
4. For the top 3, propose a specific prevention step I can take in the next two weeks.
5. Finish with one question I should answer before going ahead.
Meta-promptingAgents & reasoning
Turns a vague idea into a precise system prompt for an agent, GPT or assistant.
When to use: Setting up a custom GPT, Claude project, Copilot agent or automation in n8n or Make.
I want to build an AI agent that: [DESCRIBE WHAT IT SHOULD DO, FOR WHOM].
First ask me up to 5 short questions that would most improve the result. Wait for my answers.
Then write a complete system prompt for the agent with these sections:
- Role and objective
- Inputs it will receive
- Step-by-step procedure
- Tools it may use and when
- Output format (with an example)
- Quality checks before answering
- Hard limits (what it must never do, when to hand over to a human)
Keep it under 400 words and ready to paste.
Hypothesis-driven analysisCode & data
Explores a dataset, forms hypotheses, tests them and reports what actually matters.
When to use: Sales exports, survey results, KPIs — any table you need insights from.
You are a data analyst. Data: [PASTE CSV / TABLE OR DESCRIBE THE FILE]. Business question: [WHAT I WANT TO KNOW].
1. Profile the data: columns, types, missing values, obvious errors.
2. Propose 3–5 hypotheses that could answer the question.
3. Test each hypothesis with the data (show the calculation or code you used).
4. Report: which hypotheses hold, which don't, and how strong the evidence is.
5. Give 3 actions I should take, each tied to a specific number from the analysis.
Flag any conclusion that is based on too little data.
Adversarial reviewCode & data
Reviews code like a senior engineer hunting for real bugs, not style nitpicks.
When to use: Before merging a change or deploying code you wrote with or without AI.
Review this code as a senior engineer. Language / framework: [e.g. TypeScript, Next.js].
What it should do: [SHORT DESCRIPTION].
[PASTE CODE]
1. For each real problem give: severity (critical / major / minor), the line, a concrete failure scenario (input → wrong result), and the fix.
2. Focus on correctness, security, data loss, concurrency and edge cases. Ignore pure style.
3. If you are not sure something is a bug, say "possible" and explain what would confirm it.
4. End with the single most important change to make first.