Privacy choices: we'd like to measure visits anonymously (our own cookie-free counter plus Google Analytics, which sets its own cookies; no ads) and load Beehiiv's newsletter sign-up, which may set its own cookies. Nothing optional runs until you choose. Privacy policy
Get a written plan and a chance to correct it before the agent touches your code.
Use it when: Starting any change bigger than a one-line fix.
Show promptHide prompt
Before you change any code, read [FILES OR AREA] and write a short plan: what you will change, in what order, and what could break. List the tests you will run to prove it works. Then stop after the plan and wait for me to say go. Do not write code yet.
Review this diff like a senior engineer
Coding agentsClaude CodeCodex
A review that finds real bugs instead of praising your code.
Use it when: Before you open a pull request.
Show promptHide prompt
Review the diff below as a senior engineer who has to maintain this code. For each issue, give the file and line, what is wrong, and why it matters. Rank issues: blocker, should fix, nit. Ignore style nits unless they hide a bug. Check specifically for: [1] logic errors and off-by-one mistakes, [2] missing error handling, [3] security issues (injection, secrets, auth), [4] tests that do not actually test the change. End with a one-line verdict: ship it, fix first, or needs rework.
Diff: [PASTE DIFF OR PR LINK]
Fix the failure, not the test
Coding agentsClaude CodeCodex
Stops the agent from weakening a test just to make it green.
Use it when: A test is red and you are tempted to weaken it.
Show promptHide prompt
This test is failing: [TEST NAME AND FAILURE OUTPUT]. Find the real cause. Rules: do not change the test's assertions to make it pass, do not delete the test, and do not skip it. If the test itself is wrong, prove it: show what the code actually does versus what the test expects, and say which one is correct and why. Then fix the code (or, only with proof, fix the test) and show the test passing.
Refactor without changing behavior
Coding agentsClaude CodeCodex
Clean up code that already works, with the tests as the guardrail.
Show promptHide prompt
Refactor [FILE OR FUNCTION] to [GOAL: e.g. remove duplication / make it readable / split it up]. Do not change what it does: same inputs, same outputs, same edge cases. Keep the public interface identical. After each step, run [TEST COMMAND] and stop if anything fails. At the end, list what changed and what you deliberately left alone.
Reproduce the bug before you fix it
Coding agentsClaude CodeCodex
No fix without a repro: proves the bug is real and the fix worked.
Use it when: A bug report lands with no repro steps.
Show promptHide prompt
Bug report: [PASTE REPORT]. Before fixing anything, write the smallest reproduction you can: the exact input, the exact steps, and the wrong output you see. If you cannot reproduce it, say so and list what you tried and what information is missing — do not guess a fix. Once it reproduces reliably, fix the cause (not the symptom), re-run the reproduction to prove it is gone, and run [TEST COMMAND] to check for regressions.
MCP
Starter
Wire up an MCP server, then prove it works
MCPClaude CodeCursor
Add the server, list its tools, and smoke-test one read-only call.
Use it when: Adding a new MCP server to your agent.
Show promptHide prompt
Add the MCP server [SERVER NAME] to my agent config. Its [COMMAND / URL] is [CONNECTION DETAILS]. After wiring it up: 1) list the tools it exposes, 2) call the simplest read-only tool once to prove the connection works, 3) tell me what each tool can do and what permissions it needs. If a tool asks for credentials or write access, stop and ask me first — do not approve anything on your own.
Audit what an MCP server can touch
MCPClaude CodeCursor
A permission check before you connect a server you have not vetted.
Show promptHide prompt
Before I connect this MCP server, audit it: [SERVER NAME / REPO OR CONFIG]. Tell me: 1) every tool it exposes and whether each one reads, writes, deletes, or runs code, 2) what data it can reach (files, network, credentials, other apps), 3) anything it sends outside my machine, 4) any known issues or red flags in its repo or docs. End with a verdict: safe to connect, connect with limits (say which), or do not connect — and why.
Debug a silent MCP tool call
MCPClaude CodeCursor
A step-by-step diagnosis when a tool returns nothing, errors, or hangs.
Show promptHide prompt
This MCP tool call is failing: server [SERVER], tool [TOOL], arguments [ARGUMENTS], result [ERROR OR EMPTY RESPONSE]. Diagnose it step by step: 1) is the server running and reachable, 2) are the arguments valid for this tool's schema, 3) is it an auth or permissions problem, 4) is the server returning an error it is swallowing? Fix the smallest thing that is actually wrong. Do not retry the same call more than twice, and do not invent a successful result.
Pick the right MCP server for the job
MCPClaude CodeCursor
Compare candidates on what they really do, setup effort, and upkeep.
Use it when: Two or more MCP servers claim to do the same thing.
Show promptHide prompt
I need an MCP server for [TASK]. Candidates: [SERVER A], [SERVER B], [SERVER C]. Compare them on: 1) does it actually do [TASK] or only part of it, 2) setup effort and what it needs from me, 3) what it can read, write, or send, 4) maintenance — last update, open issues, who runs it. Recommend one, name the runner-up, and say what would change your pick.
Write a minimal MCP server spec
MCPClaude CodeCursor
A tight spec for a small custom server: tools, inputs, and limits only.
Use it when: You need a small custom MCP server and want it built once, properly.
Show promptHide prompt
Design a minimal MCP server for [PURPOSE]. Keep it to at most [NUMBER] tools. For each tool, give: name, what it does, inputs (name, type, required or optional), outputs, and what it must never do. List the permissions it needs and nothing more. Then write the config snippet to add it to [AGENT: Claude Code / Cursor / ...]. Do not add features I did not ask for — if something is nice-to-have, list it separately as 'later'.
Browser agents
Starter
Browse, but cite every claim
Browser agentsBrowser UseOpenAI Operator
Research where a wrong answer costs you: every claim links to its source.
Show promptHide prompt
Research [QUESTION] on the web. Rules: 1) every factual claim in your answer must link to the page it came from, 2) prefer primary sources (docs, official announcements, the data itself) over roundups, 3) if sources disagree, say so and show both, 4) if you cannot verify something, mark it unverified — do not present it as fact. End with a short list of the sources you actually used.
Fill this form without guessing
Browser agentsBrowser UseOpenAI Operator
Form automation with a hard rule: missing values stay blank, nothing submits.
Use it when: An agent has to complete a form with your real details.
Show promptHide prompt
Fill in the form at [URL] using only these details: [DETAILS]. Rules: never guess a field — if a value is missing, leave the field blank and list it at the end. Never submit the form; stop at the review screen and show me every field before anything is sent. Never enter payment details, passwords, or anything I did not list here.
Compare prices across three stores
Browser agentsBrowser UseOpenAI Operator
A price check with evidence: price, shipping, stock, and a link for each store.
Use it when: Buying something and the price varies by store.
Show promptHide prompt
Compare the price of [PRODUCT, EXACT MODEL] at [STORE A], [STORE B], and [STORE C]. For each store give: the price, shipping, delivery time, and whether it is in stock — each with a link to the page you checked. Flag any listing that looks like a different model, a refurb, or a marketplace seller. End with the cheapest real total and what would change the answer.
Watch this page and tell me when it changes
Browser agentsBrowser UseOpenAI Operator
A quiet watcher: only speaks up when the thing you care about actually changes.
Use it when: Waiting on a price drop, restock, or announcement.
Show promptHide prompt
Check [URL] once a day and tell me only when [WHAT TO WATCH: e.g. the price drops below $X / it is back in stock / the announcement appears]. Each check, record the date and what you saw in one line. Do not message me when nothing changed. When it changes, send me the new value, a link to the page, and a screenshot or quote of the changed part.
Summarize this page for a decision
Browser agentsBrowser UseOpenAI Operator
Turns a long page into the facts, the cost, the catch, and a recommendation.
Use it when: A long page stands between you and a yes/no.
Show promptHide prompt
Read [URL] and summarize it for this decision: [DECISION, e.g. should I buy this / sign up for this / switch to this]. Give me: 1) what it is, in two sentences, 2) the three facts that matter most for my decision, 3) the cost and the catch, 4) what it does not tell me that I would need to know. End with your recommendation and one line on why — and mark anything you could not verify.
Have a prompt worth sharing?
Send us one prompt that makes a coding, MCP or browser agent do better work. Hand-written and original only — we check every prompt by hand before it goes live.