one-minute-ci
Make a repo's CI fast without letting it lie. Audits the last 150 pull request runs, measures PR wall-clock at p50 and p95 against a target of under 1 minute and under 2, splits every job into blocked, waiting, setup and running time, and fixes the biggest bucket first, one small PR per change with before and after numbers from real runs. Holds a budget per stage (runner pickup, setup, lint, format, typecheck, tests per shard, local hooks) and fails a job that goes over. Refuses fake speed-ups - retries, wider timeouts, skipped tests, `|| true`, `paths-ignore`, a shorter required-check list - and keeps the full suite in a merge queue before main. Defaults to Bun (`bun test`, `bun check`, `bun fmt`, `bun lint`), Blacksmith runners and hk hooks, each proven side by side before the switch. Use when the user says "speed up CI", "CI is slow", "make CI faster", "get CI under a minute", "why does my PR take 10 minutes", "audit our GitHub Actions", "where does our CI time go", "report CI time and cost", "set up a merge queue", "shard the tests", "the tests are flaky, just add a retry", "add pre-commit and pre-push hooks", or "move our CI to Bun".
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/kleenpulse--one-minute-ci)HTML
<a href="/listings/kleenpulse--one-minute-ci"><img src="/api/listings/kleenpulse--one-minute-ci/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
npx skills add kleenpulse/one-minute-ci
Works with Claude Code, Codex, Cursor, Copilot, and anything else that reads `SKILL.md`. The audit script needs [Bun](https://bun.sh) and an authenticated [`gh`](https://cli.github.com).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 170 lines
- ---
- name: one-minute-ci
- description: Make a repo's CI fast without letting it lie. Audits the last 150 pull request runs, measures PR wall-clock at p50 and p95 against a target of under 1 minute and under 2, splits every job into blocked, waiting, setup and running time, and fixes the biggest bucket first, one small PR per change with before and after numbers from real runs. Holds a budget per stage (runner pickup, setup, lint, format, typecheck, tests per shard, local hooks) and fails a job that goes over. Refuses fake speed-ups - retries, wider timeouts, skipped tests, `|| true`, `paths-ignore`, a shorter required-check list - and keeps the full suite in a merge queue before main. Defaults to Bun (`bun test`, `bun check`, `bun fmt`, `bun lint`), Blacksmith runners and hk hooks, each proven side by side before the switch. Use when the user says "speed up CI", "CI is slow", "make CI faster", "get CI under a minute", "why does my PR take 10 minutes", "audit our GitHub Actions", "where does our CI time go", "report CI time and cost", "set up a merge queue", "shard the tests", "the tests are flaky, just add a retry", "add pre-commit and pre-push hooks", or "move our CI to Bun".
- license: MIT
- ---
- # One-Minute CI: Fast, and It Never Lies
- You make a repo's CI as fast as it can go while a green check keeps meaning what it meant before. The target is a pull request that gets its result in under 1 minute at the median and under 2 minutes at p95. You measure first, fix the biggest cost, and prove every change with numbers from real runs.
- The target, the three rules that outrank speed, the budgets and the tool defaults come from a prompt by Pontus Abrahamsson: ["Updated: prompt to get your CI under 1 minute."](https://x.com/pontusab/status/2123771160261529982) (20 Nov 2026). The audit script, the proofs, the recipes, the self-check and the rules marked *(added)* are this skill's own. This skill is not affiliated with or endorsed by Pontus Abrahamsson, Blacksmith or hk.
- ## When to Use
- - CI is slow and someone wants it faster
- - Someone asks where the CI time goes, or what CI costs
- - A repo is moving its lint, format, typecheck or tests to Bun
- - Someone wants a merge queue, test shards, test selection, or local hooks
- - Someone wants to make a slow or flaky check pass by retrying it, skipping it or raising a timeout (use this skill to say no and find the cause)
- **Do not use for:** fixing one failing test, deploy and release pipelines, CI that is not GitHub Actions (the audit script reads the Actions API; the rules still apply), or getting to green by checking less.
- ## Core Principle
- **Speed only counts if a green check still means what it meant before.** Anyone can make CI fast by running less of it. Every change here must leave the repo checking the same things or more, and must show its saving in real run times.
- ## Rules
- ### Three rules that outrank speed
- 1. **The full suite runs before anything reaches main, in a merge queue.** Test selection only narrows the feedback on a pull request. It never decides what merges.
- 2. **Never hide failures.** No retries, no wider timeouts, no skipped tests. A check that needs a retry has a bug. Find it.
- 3. **Never check less.** Lint rules and required checks only stay or grow.
- ### Rules for the work
- 4. **Audit before you change anything.** Run `references/audit.ts` and fix the biggest bucket first. No audit, no change.
- 5. **Real numbers only.** Before and after numbers come from real runs, never from an estimate. If you have no "after" yet, say so.
- 6. **One small PR per change.** Each PR carries its before and after. Two changes in one PR means nobody knows which one helped.
- 7. **A change to a judge goes alone** *(added)*. Deleting tests and swapping a test runner or a linter both change what "green" means. Each goes in its own PR that a human approves, never inside a speed-up PR.
- 8. **Side by side before every switch.** Run the old tool and the new tool on the same commit and list everything the new one misses. The tools below are defaults; keeping an old one needs a measured reason, and so does dropping it.
- 9. **Prove the built-in is what runs** *(added)*. Find the command in `bun --help` on the pinned version, and use `--no-script`. "It ran and exited 0" is not proof: with a `lint`, `fmt` or `check` script in `package.json`, `bun lint`, `bun fmt` and `bun check` run that script instead, and CI reports a check that never happened. Details in `references/recipes.md`, section 1.
- 10. **A security gate stays** *(added)*. A `needs:` that puts a security scan (a malware tripwire, a forged-commit check, a secret scan) ahead of install or code execution is not a speed problem to remove. Report what it costs in seconds and ask. This outranks "no required job waits on another".
- 11. **A budget failure is a bug.** A job that goes over its budget fails. Fix the cause. Never re-run it to get green, and never raise the budget in the PR that broke it.
- 12. **Ask once, with a price.** Do everything you can without the human first. Then bring everything that needs them as one list (see below).
- 13. **Stop at the floor** *(added)*. If the target can't be reached without breaking rule 1, 2 or 3, report how fast it got, with real numbers, and what stands in the way. Then stop.
- ### Budgets
- | Stage | Budget | Default way to get there |
- | --- | --- | --- |
- | Waiting for a runner | under 5 s | Move every job off GitHub's hosted runners to Blacksmith and keep machines warm. Cancel a branch's in-progress run when a new commit lands. |
- | Setup | under 15 s | Bake the toolchain into the runner image. Get PR cache hits above 95% by seeding caches from main, and cache `node_modules` itself. Use a blobless clone (`filter: blob:none`). |
- | Slowest job | under 35 s | No required job waits on another (but see rule 10). |
- | Lint | under 1 s | `bun lint`. It reads the existing `.oxlintrc.json`. Turn on every rule you can. |
- | Format | under 1 s | `bun fmt`. |
- | Typecheck | under 5 s | `bun check` on TypeScript 7. |
- | Tests | under 30 s per shard | `bun test`. Test files keep their `vitest` imports. Isolate modules, not processes, reset state after every file, and run in random order to flush out leaks. |
- | Local hooks | under 9 s | hk hooks. Pre-commit formats and lints staged files; pre-push runs the same check as CI. |
- The Actions API reports whole seconds. Budgets under 5 seconds are timed inside the step, not read from the API (`references/recipes.md`, section 7).
- Checked on 21 Nov 2026: `bun check` and `bun fmt` are in stable Bun 1.4.5. `bun lint` is only in the canary (1.4.6-canary.1), where it does read an existing `.oxlintrc.json`. Whether `bun check` matches TypeScript 7 was not checked by this skill; that line is the prompt's. Versions move, so apply rule 9 on the version the repo pins.
- ### The rest of the prompt's rules
- | Topic | Rule |
- | --- | --- |
- | Install and scripts | Bun. Delete npm, yarn, pnpm, Prettier, ESLint, oxlint, Vitest and Jest. If a tool only exists in a canary, pin that exact canary. |
- | Shards | Buy cores before shards. Match workers to cores. Give each shard its own throwaway database. Split any test file over 20 seconds. Balance shards from file times that main records on every full run, and prove the shards together ran every file. |
- | Fewer tests, on evidence | For two weeks, log every test failure you fix as a caught regression or a change detector. Delete the files that only detect change and ban their patterns in AGENTS.md. Never delete tests for authorization, privacy, migrations or concurrency. |
- | Shape | Run lint, typecheck, build and tests in parallel. Fold the small checks into one job that lists every failure. |
- | Selection | On PRs, run only the tests the diff can reach, and list the inputs that rerun everything. Then hide a real change from it on purpose and confirm CI goes red. |
- | Local | A hook, not a line in AGENTS.md, because agents skip steps. |
- | Keep it fast | Fail any job that exceeds its budget. |
- ### What only the human can do
- Bring these as one list, each with what it unlocks and its expected monthly cost:
- - A new account or a GitHub app (for example Blacksmith)
- - A secret
- - A bigger bill: more cores, more runner minutes, more cache storage
- - Repo settings: the merge queue, rulesets, the list of required checks
- - Deleting tests (with the two weeks of evidence)
- - A `needs:` chain that is a security gate
- For cost, use the runner's published price per minute and say which price you used. If you can't confirm a price, give the minutes and say the cost is not calculated.
- ## What to Keep (Untouchables)
- - The list of required checks and every lint rule: never shorter, never weaker.
- - Test assertions: never loosened to make a run pass.
- - Tests for authorization, privacy, migrations and concurrency: never deleted, whatever the evidence log says.
- - Security scans and the `needs:` that puts them first.
- - Run times: reported exactly as measured, with the run they came from.
- - A failing or slow run: never re-run to replace it with a better one.
- ## Method
- 1. **Audit.** From the repo, run:
- ```sh
- bun <skill folder>/references/audit.ts # the current repo
- bun <skill folder>/references/audit.ts owner/repo # or a named one
- ```
- It needs Bun and an authenticated `gh`, and it only reads. It pulls the last 150 pull request runs and prints:
- - **PR wall-clock** per commit, p50 and p95, against the target. Commits with a cancelled, unfinished or re-run run are dropped and counted, and failed commits are timed apart, so the number can't flatter the repo.
- - **Per workflow**, so an optional workflow can be told from the required one.
- - **Four buckets per job:** blocked (behind another job), waiting (for a runner), setup, running.
- - **The step mapping,** showing how every step was counted, with a list of steps it could not classify.
- - **Checks that can't fail the job today** (`|| true`, `continue-on-error`) and **jobs that wait on another**.
- - **Runner minutes,** for the cost line.
- Read the mapping before trusting the buckets. If a step is in the wrong bucket, fix it with `--map` and run again.
- 2. **Pick the biggest bucket.** One change that addresses it. State the bucket, the number, and the rule or budget it serves.
- 3. **If the change swaps a tool, prove it side by side first.** Runner swap: same test files collected and same test count, old against new. Lint swap: the new findings cover the old findings on the current tree, plus seeded violations. Recipes 12 and 13.
- 4. **Open one small PR.** In the description: the bucket, the before number with its run links, what was changed, and what was not changed (required checks, lint rules, tests).
- 5. **Measure after.** Run the audit again once the PR has real runs. Put the after number in the PR. If it didn't help, say so and revert.
- 6. **Repeat from step 2** until the target is met or rule 13 stops you.
- 7. **Hand over the list for the human** as soon as you hit the first item only they can do. Keep working on everything that doesn't depend on it.
- 8. **Finish with the totals:** PR wall-clock p50 and p95, and runner minutes and cost, before and after, each from a real audit run.
- ## Worked Example
- **Request:** "Our CI is slow. Get it under a minute."
- **Without the skill:**
- > Added a retry to the flaky e2e test, made lint a warning so it stops blocking, and added `paths-ignore` for docs. CI now finishes in about 1m 40s.
- **With the skill** (numbers from a real audit of a TypeScript repo on Bun):
- > **Audit, 150 PR runs = 139 commits.** PR wall-clock on 110 green commits: p50 2m 28s, p95 3m 31s. Target missed on both. 23 commits dropped (22 cancelled, 1 unfinished). 6 failed commits, timed apart.
- >
- > **Biggest bucket:** `check`, running, p50 1m 57s. Its steps run one after another in one job: lint 46 s, build 38 s, tests 15 s, typecheck 14 s. Waiting is 3 s and setup is 10 s, both inside budget.
- >
- > **Found on the way:**
- > - Lint runs as `bun run lint || echo ...`. It takes 46 s and cannot fail the job. I did not remove it. Making it a real gate is on the list below.
- > - `check` needs `scan`, and the workflow says nothing is installed before the scan passes. That is a security gate. It costs about 11 s at p50. I kept it.
- >
- > **PR 1:** run lint, typecheck, tests and build as parallel jobs, each still behind the scan. No check removed, no rule changed. Before: p50 2m 28s (runs linked). After: to be filled in from real runs once this lands.
- >
- > **Needs you:**
- > 1. Blacksmith account and GitHub app, to cut runner pickup and setup. Cost: not calculated yet; the audit shows 3.6 billed minutes per commit today, and I need your plan's price per minute.
- > 2. Merge queue and required checks, in repo settings. No cost.
- > 3. Make lint a gate? It reports problems today that nobody is forced to fix.
- > 4. Keep the scan in front of everything (about 11 s per PR)? I recommend keeping it.
- >
- > **Refused:** a retry for the e2e test. A test that needs a retry has a bug; I'll find it in a separate PR.
- The first answer is faster and proves nothing: three checks now mean less than before. The second answer has not saved a second yet, but every number is real, nothing checks less, and the human knows exactly what is theirs to decide.
- ## Self-Check Before You Deliver
- 1. The report opens with PR wall-clock p50 and p95 from `audit.ts`, with the number of commits, and the number dropped and why.
- 2. Every "before" and "after" names the runs it came from. No number is an estimate presented as a measurement.
- 3. The diff adds none of these (search for each): `retry`, `continue-on-error`, `.skip`, `.only`, `xit`, `if: always()`, `paths-ignore`, `--pass-with-no-tests`, `--allow-unsupported`, a larger `timeout-minutes`, a larger `--timeout`.
- 4. Every `||` after a check command in the diff carries the failure to a non-zero exit. `|| true`, `|| echo` and `|| exit 0` fail this check.
- 5. The list of required checks is the same length or longer. No lint rule was turned off.
- 6. No `needs:` in front of a security scan was removed or bypassed.
- 7. No test was deleted in a speed-up PR. Any deleted test has its own approved PR and two weeks of evidence, and is not about authorization, privacy, migrations or concurrency.
- 8. Every Bun built-in in the diff was found in `bun --help` on the pinned version and runs with `--no-script`, or the old script was deleted in the same PR.
- 9. Each tool switch has its side-by-side list in the PR.
- 10. If shards were added: the shard proof ran and every test file was run by one shard.
- 11. If selection was added: a hidden real change was pushed and the PR run went red. The run is linked.
- 12. Each budget is enforced by a step that fails, and the sub-5-second ones are timed inside the step.
- 13. Everything that needs the human is in one list, each item with a cost or the words "not calculated".
- 14. If the target was not met, the report says how fast it got and which rule stopped it.
README.md · README · 76 lines
- # one-minute-ci
- Makes a coding agent speed up your CI without letting it lie: a pull request gets its result in under 1 minute, and a green check still means what it meant before.
- ## Install
- ```bash
- npx skills add kleenpulse/one-minute-ci
- ```
- Works with Claude Code, Codex, Cursor, Copilot, and anything else that reads `SKILL.md`. The audit script needs [Bun](https://bun.sh) and an authenticated [`gh`](https://cli.github.com).
- ## Credit
- The target, the three rules that outrank speed, the budgets and the tool defaults come from a prompt by **Pontus Abrahamsson** ([@pontusab](https://x.com/pontusab)): ["Updated: prompt to get your CI under 1 minute."](https://x.com/pontusab/status/2123771160261529982) (20 Nov 2026).
- This skill turns that prompt into something an agent can run and be checked against. The audit script, the proofs, the recipes and the self-check are this skill's own additions, and so are the rules marked *(added)* in `SKILL.md`. It is not affiliated with or endorsed by Pontus Abrahamsson, Blacksmith or hk.
- ## Why this one
- Ask an agent to "make CI faster" and the quickest path is to check less: add a retry, skip the slow test, turn lint into a warning, ignore some paths. The pipeline gets faster and the green check gets weaker, and nobody notices until a bug ships.
- This skill makes that path unavailable. It measures first, with a script that can't flatter the repo. It fixes the biggest cost, one small PR at a time, with before and after numbers from real runs. And it has a searchable list of the shortcuts it will not take.
- ## What it does
- - **Audits the last 150 pull request runs** with `references/audit.ts`: PR wall-clock at p50 and p95 against a target of under 1 minute and under 2, then four buckets per job (blocked behind another job, waiting for a runner, setup, running).
- - **Keeps the headline honest:** commits with a cancelled, unfinished or re-run run are dropped and counted, and failed commits are timed apart, since a failure often arrives sooner than a pass.
- - **Shows its work:** prints how every step was classified, lists the ones it could not classify, and names the checks that can't fail the job today (`|| true`, `continue-on-error`).
- - **Holds a budget per stage:** runner pickup under 5 s, setup under 15 s, slowest job under 35 s, lint and format under 1 s each, typecheck under 5 s, tests under 30 s per shard, local hooks under 9 s. A job that goes over fails.
- - **Moves the toolchain to Bun** (`bun test`, `bun check`, `bun fmt`, `bun lint`), runners to Blacksmith and local hooks to hk, each proven side by side with the old tool before the switch.
- - **Proves the shards ran everything** with `references/shard-proof.ts`, and proves test selection by hiding a real change from it and watching CI go red.
- - **Brings you one list** of what only you can do (accounts, secrets, bigger bills, repo settings), each with its expected monthly cost.
- - **Ends with totals:** CI time and cost, before and after, from real runs.
- ## What it refuses to do
- - Add a retry, a wider timeout or a skipped test to make a run pass
- - Turn off a lint rule or shorten the list of required checks
- - Add `|| true`, `continue-on-error`, `paths-ignore` or `if: always()` to get to green
- - Let test selection decide what merges; the full suite runs in a merge queue before main
- - Delete a test inside a speed-up PR, or ever delete tests for authorization, privacy, migrations or concurrency
- - Remove a `needs:` that puts a security scan ahead of install or code execution
- - Re-run a slow or failing job to replace it with a better one
- - Trust `bun lint`, `bun fmt` or `bun check` because "it ran": a `package.json` script with the same name takes over silently, so it checks `bun --help` and uses `--no-script`
- - Report an estimate as a measurement
- - Keep going past the point where the target needs one of its rules broken; it reports how fast it got and stops
- ## Usage
- - "Our CI is slow. Get it under a minute."
- - "Audit our GitHub Actions and tell me where the time goes"
- - "Why does my PR take 10 minutes to go green?"
- - "Move our lint, format and typecheck to Bun"
- - "Shard the tests and prove nothing got dropped"
- - "The e2e test is flaky, just add a retry" (it will refuse, and find the cause)
- Run the audit on its own, without an agent:
- ```bash
- bun references/audit.ts owner/repo
- bun references/audit.ts owner/repo --runs 150 --json
- ```
- ## What was checked
- Every command in `references/recipes.md` was run, or read in the tool's own help or official docs, on 21 Nov 2026, and the file says which version. Two things worth knowing before you start:
- - On Bun 1.4.5 (stable), `bun check` and `bun fmt` are built in. `bun lint` is only in the canary (1.4.6-canary.1 at the time). The skill makes the agent confirm this on the version your repo pins.
- - Blacksmith's warm-machine and custom-image settings, hk's `pre-push` syntax, and `bun check`'s relation to TypeScript 7 were not checked here. The skill states those as rules in words and points to the tools' own docs.
- The audit script has unit tests: `bun test references/`.
- ## License
- MIT
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.