Press / to search here, or ⌘K anywhere on the site to copy a prompt without leaving the page.
Start from the job, not the list
These are the jobs people actually arrive with. Run the steps in order; where a step has a fork, take the branch that matches your situation.
-
Turn a vague ask into a change you can defend
A one-line request, and every expensive decision still unmade.
-
Change old code without breaking it
No tests, real users, and something that has to get faster or cleaner.
-
Let an agent touch something you cannot undo
A delete, a migration, a force-push, a deploy.
-
Run an agent overnight and wake up to proof
You will not be watching, and optimistic prose is not evidence.
-
Decide whether to believe a green result
Everything passed, and you cannot say what any of it proved.
-
Ship a screen that does not feel machine-made
It builds, it ships, and it still feels like a machine made it.
-
Recover stranded work and leave the repo clean
Stashes, stale branches, orphaned worktrees, and a pile of open PRs.
-
Write something that reads human and survives a fact-check
Technically right, and it still sounds like a model wrote it.
The full index
30 prompts, each with receipts
-
Divergence gate
Explore, don’t decide: widen the field before the model converges
Use in the moment before you ask a model to recommend, so the first plausible answer cannot anchor you.
-
Onboarding
Map an unfamiliar repo from evidence, not guesses
Use on day one in a repo you have never seen and cannot yet trust anyone's summary of.
-
Refactor
Refactor under a net – pin the behaviour, then change it
Use before refactoring code that has no tests and real users.
-
Spec
Turn a vague request into a spec, with every assumption tagged
Use when the request is one line and the expensive decisions are all unstated.
-
Performance
Measure first – find the real bottleneck, keep only proven wins
Use when something is slow and everyone already has a theory about why.
-
Debugging
No fix until the bug reproduces
Use when an agent keeps offering fixes for a bug nobody has reproduced.
-
Git workflow
Git hygiene and work-recovery sweep
Use when real work is stranded in stashes, stale branches, and orphaned worktrees.
-
Fix verification
Prove the bug is gone, not just that CI went green
Use when CI went green after a fix and that is all you know.
-
QA audit
Vacuous-pass audit: find the checks that pass without checking
Use when everything is green and you cannot say what any of it actually proved.
-
Sign-off gate
Sign-off gate: prove the claim or label it NOT RUN
Use when an agent says done, passing, or safe to deploy, and you want the receipt.
-
Review synthesis
Two adversarial critics, you stay the judge (dedupe + keep uniques)
Use when two models have reviewed the same work and you have to be the judge.
-
QA review
Multi-lens QA fan-out: five blind reviews, then triage
Use when one review pass keeps rationalising away what another pass would catch.
-
Verify before encode
Claim-ladder gate: verify against the live source before you encode it
Use before you write a remembered fact into a config, a guardrail, or a doc.
-
Integrity auditexperimental
Audit a byte-exact copy-paste claim end to end
Use when a copy button, an export, or a docs snippet hands people something subtly different from the source.
-
Code review
A code reviewer that refuses to rubber-stamp
Use when reviews keep coming back approved with nothing said.
-
PR review
Cherry-pick the value from every PR (never merge dirty)
Use when open PRs have piled up and each has one good part inside a mess.
-
Optimisation
Audit your governance, then level it up (enforced, not theatre)
Use when a project has rules and gates and nobody has checked whether any of them fire.
-
Abstain
Abstain when the data is stale: suppress, do not guess
Use when something reports confidently on data that may be stale, thin, or missing.
-
Data safety
Plan, then apply – a two-phase gate for irreversible actions
Use before anything you cannot undo: a delete, a migration, a force-push, a deploy.
-
Audit guard
Audit guard: stop agents fixing work you deliberately deferred
Use before any quick-wins or tech-debt sweep in a repo with documented deferrals.
-
Safety audit
Fail-open or fail-closed: audit a gate before you approve it
Use when reviewing a guard and nobody can say what it does when it errors.
-
Dispatch contract
Builder-lane dispatch: brief a second agent, judge its work
Use when you are handing real code to a second agent and you will judge the result.
-
Agent dispatch
Brief a sub-agent like a colleague, not a warrant
Use when a sub-agent follows your instructions to the letter and still misses the point.
-
Agent ops
Unattended autonomous work orchestrator: validate, then implement
Use before a long run you will not be watching.
-
Session handoff
Session handoff and durable maintenance, double-audited
Use at the end of a session a stranger has to pick up cold tomorrow.
-
Agent ops
Keep a long agent run from rotting
Use when a run is long enough that the agent's own context is the thing degrading.
-
Review pass
Think Like Roy review pass: score work against a standard it must cite
Use when a screen passes review and still feels generic, and you want the objection named.
-
UI review
AI-feel block audit: five laws before a screen ships
Use when a screen looks fine block by block and reads AI-made as a whole.
-
Writing
Sharp engineering writer: four self-critiques that kill AI tells
Use when the draft is technically right and reads like a model wrote it.
-
Product/UX review
Module-by-module product and UX review
Use when the product works and you have not walked it as a user in months.
Nothing matches that. Try the jobs above, or read the index at /prompts.json.
Shared as is, for reference. Read a prompt and decide what it will do before you run it; using it is your responsibility, under the terms.