agentleFS
FracteraFractera/Agent-Engineering-Infrastructure/.claude/skills/use-testing/SKILL.md

use-testing

Skill61 starsChanged 55 days agoMITView on GitHub

Build process. How work is PROVEN here — before you say «done», «it works», «ready to check». Load it when you are about to report a result, when a gate goes green and you feel finished, when something must be verified on the node, and when the owner asks for proof («show me the proof», «докажи»). The rule that changes behaviour: two proofs from DIFFERENT planes, one of them with a negative control, and a build log is never one of them — it reads the same whether the capability works or not.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every guard passed before you touched anything.

    line 27
  2. Distinctive:

    No key, the owner's session, a phone, a Telegram account — name the missing proof in the report, in the same message, and do not replace it …

    line 59
  3. Distinctive:

    If point 1 cannot be filled, the work is not finished: it is unmeasured.

    line 91

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

use-testing

Process: build. A hint, not a law, in HOW you prove: know a better way for the case in front of you — use it and say so.

1. Two proofs, from different planes

A step or a substep closes on two proofs from different planes — otherwise the same thing was measured twice.

PlaneAnswers
the treeis the code well-formed — guards, types, structure
the wirewhat the node actually answers — status, headers, body of a door or a page
the eyewhat a person sees — a real browser, both themes, a narrow screen, a real mouse
the datawhat was written — a row, a file, a commit, a tag on GitHub
the storyis this the thing that was ASKED for — the owner's word

A proof has four fields: what was run · the verbatim output · what it proves · how it would look without the change. The last field separates a proof from a ritual. A build log is never one of the two: «it compiles», «200», «RC=0» say nothing of the capability.

2. Green was green BEFORE your change — so one proof carries a negative control

Every guard passed before you touched anything. So one proof breaks the thing on purpose, sees it fail, and puts it back: spoil the input (a wrong name, a missing section, a returned row), remove the fix (old code, same probe), or ask about something known to be absent. A guard that stays green on a spoiled input is blind — and that is a finding, not a pass. When the live data cannot be changed without the owner (a published element, a person's records), spoil the input on the way instead: a copy in a temporary folder, or the door's answer replaced inside the browser (CDP Fetch.fulfillRequest).

3. The defects no guard catches

Markup can be correct and wrong: a page that reads like another site, blocks that do not tell a story, a section repeated in another guise. The proof is comparison — the page and its neighbour side by side — and the owner's eye. Scrolling, timing, anything a page does by itself is proven in a real browser with a person's timing, not by a fixed pause.

4. Measurements that lied here — each looked like success

  • /api/health → commit is the commit the process started on; builtAt is the build. After a rebuild also count the page's /_next/static/* files that do not answer 200.
  • npm run build | tail prints tail's exit code, always 0. Log to a file, RC=$? right after npm, require a marker.
  • pm2 online for a process in an endless restart loop: watch the restart count, uptime GROWING between two samples, and who holds the port.
  • A page whose data an island fetches after load has none of that data in its server HTML — a curl finds nothing even when it works. Prove it in a browser, or ask the door the island asks.
  • A search over a whole page finds the navigation: menus, crumbs and footers print every page's title. Address the element that belongs to this page (<h1>, a data-* mark); the same check on ANOTHER page must answer differently.
  • git status under core.autocrlf shows «M» for a file rewritten with the same text; ask git diff --cached HEAD after git add.
  • Search by meaning always returns the nearest neighbour — «nothing found» needs a threshold or a filter.
  • grep in Git Bash: -P fails on the locale and reads as 0, -i does not fold Cyrillic, -c $'\r' counts every line. Zero is not «not there» until a known-present item gives one.
  • The service worker serves yesterday's build in today's browser: unregister it AND clear caches.
  • A hidden browser tab reports width 0; check document.visibilityState before trusting an automated measurement.

5. What cannot be proven — say it, never substitute

No key, the owner's session, a phone, a Telegram account — name the missing proof in the report, in the same message, and do not replace it with something easy to obtain. Never simulate what the owner does in five seconds (a voice note, a sign-in, a payment): the simulation checks another path. When the proof needs a live run of an agent with a model, it runs only by the owner's word, with its price named.

6. The form — and the owner's command

The owner may ask for proof at any moment («show me the proof», «докажи», «покажи доказательства» — one request). Answer with what is done by that minute, in this form — an empty field is visible, prose hides it:

PROOF 1 — plane <which>
  ran:          <command / request / action>
  output:       <verbatim>
  proves:       <which line means it works>
  without it:   <how the same output would look without the change>
PROOF 2 — plane <another>
  …
NEGATIVE CONTROL
  case:         <the input whose answer must differ>
  answer:       <and it differed — like this>
DELIVERY (not a proof)
  build / release / restart: <exit codes, hashes, tags>
NOT PROVEN
  <what needs the owner or is out of reach, and why>

7. Before you say «done»

  1. Name the two planes and paste the verbatim output of each.
  2. Say what would have looked different without the change.
  3. Say what remains unproven — plainly, in the same message.

If point 1 cannot be filled, the work is not finished: it is unmeasured.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.