Index of every page on this site: llms.txt
All skills
raphaelsalaja avatar

/verify

@b2e4283
Run it yourself

Checks a change in the running app before calling it done, then says where to see it.

Use this Skill: https://skilld.dev/gh/raphaelsalaja/skills/verify

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈23 tokens for metadata: the name and description. ≈1.6k when used: this file. ≈24 more on demand in 1 file.

Description uses 1.2% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈94 characters, or 1.2%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

Verify

This skill decides when a change may be called done. It asks for evidence from the running result, not from a passing build.

A build is not evidence

A typecheck, a lint pass or a green build does not show that a visual or behavioural change works. Saying "fixed" without opening the result is a finding. Opening it and reporting an imperfect result is not.

Run it after any change to UI or behaviour, before replying "done". Run it at once when the user says "still", "nothing changed" or "test it yourself". What to change belongs to scope. Matching an outside app belongs to match. Evidence behind a review finding belongs to audit.

1. See the problem first

Open the app, the page or the output and find the symptom before editing. Use the browser, a screenshot or computer use. Editing from the description alone is verify-reproduce-first. When you cannot reach the running result, say what blocks you.

2. Look where the user will look

First confirm the running app serves your edit. The dev server must run from this checkout, on the port you open, restarted after any config or .env change. lsof -iTCP -sTCP:LISTEN -n -P names the process on each port, and lsof -a -p <pid> -d cwd names its folder. Judging a stale server or another worktree is verify-served-the-edit.

Reload, or confirm hot reload applied, then look at the same place. A build pass offered as proof is verify-build-is-not-proof. Click every control you added once. A control with no effect is verify-click-new-controls.

Look at the whole surface, not only the line you edited. Scroll to both ends and open each state the change touches. Checking only the edited element is verify-whole-surface.

Read the console before and after. Done means no error or warning the change introduced, such as a missing key warning. A new console error or warning left behind is verify-clean-console.

Touch behaviour is confirmed on a real phone over remote debugging, ideally one a few years old. That covers sticky hover, tap delay, rubber-banding, safe areas and the on-screen keyboard. Check once with the keyboard open and once in landscape. Without a phone, report Not verified on device. A touch fix called done from desktop emulation is verify-on-device.

3. Keep every check honest

A check made to pass is not a pass. Never delete or skip a test, or add .skip, .only, @ts-ignore or eslint-disable to get past one. Never change an expected value to match new output, and never special-case the input a test uses. When a test contradicts the request, stop and name the conflict. Any of these is verify-no-weakened-checks.

Search your own diff for them before reporting: git diff -U0 | rg '^\+.*(\.skip\(|\.only\(|@ts-ignore|eslint-disable)'. Each hit needs a reason the user agreed to.

4. Find the cause on a repeat report

When the user reports the same problem again, stop adjusting the same value. Diff against the last commit where it worked, read the computed style or search for the cause. A third guess at one property is verify-root-cause-on-repeat.

5. Report what you saw

Say what you checked, how and where the user can see it: the route, the window, the menu path or the folder. "It should work now" is verify-say-where. Open the result for the user when you can. A bare output path is verify-open-the-result.

Attach the evidence. A visual change carries the screenshot taken after the fix. A behaviour change carries the steps and what happened. "Verified, looks good" with nothing attached is verify-show-evidence.

Name anything you could not check and why. "Done" with an unchecked item is verify-name-the-gap.

The pair

Every check is a pair. The broken line sits beside the line that replaces it.

Rule Broken Replacement
verify-reproduce-first Editing from the bug report alone Open the screen, see the bug, then edit
verify-served-the-edit Checked on port 3000 while the edit lives in another worktree The server's folder confirmed with lsof, then checked
verify-build-is-not-proof "Typecheck passes, so it is fixed" "Reloaded Settings and the toggle now persists"
verify-click-new-controls A new "Add Condition" button with no handler Click it once and see a row appear
verify-whole-surface Checking the edited row only The whole panel, scrolled to both ends
verify-clean-console "Done" with a new missing key warning in the console The warning fixed, the console as clean as before
verify-on-device A safe-area fix checked in desktop emulation Checked on a phone with the keyboard open and in landscape, or Not verified on device
verify-no-weakened-checks expect(total).toBe(42) changed to toBe(41) so the suite passes The code fixed, or "The test expects 42 and the request says 41. Which is right?"
verify-root-cause-on-repeat A third tweak to the same padding Diff against the last working commit, then fix
verify-say-where "It should work now" "Open Library, right-click a save, choose Hide"
verify-open-the-result ~/Downloads/clips printed in the reply The folder opened for the user
verify-show-evidence "Verified, looks good" The screenshot of Settings after the fix, and "clicked Save, reloaded, the value stayed"
verify-name-the-gap "Done" while export was never run "Not verified: export, it needs a signed build"

Reporting

Severity. HIGH is "done" with no look at the running result, or a new control that does nothing. It is also a check weakened to pass, or a result judged on a server that was not serving the edit. MEDIUM is an edit made before seeing the bug, a repeat report met with another guess or only the edited line checked. It is also a reply with no place to see the change, or "done" hiding an unchecked item. So is a console error or warning the change introduced, or a touch change neither tried on a phone nor marked. LOW is a checked change whose reply attaches no evidence, or a bare output path never opened.

Format. One finding per root cause, citing the request when no file applies:

file:line - [rule-id] what is wrong
request 2 - [rule-id] what is wrong

Then a Rule | Count | Severity summary. Block when any HIGH remains, Approve otherwise. Never Approve a change you did not open. Mark every check you could not run as Not verified.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 warningtoday3 checks · Risk SAFE
  • Gen Agent Trust Hubtoday

    This skill provides guidelines for verifying application changes by inspecting the running software. It uses standard tools like git and lsof to check the development environment. The skill introduces a low risk of indirect prompt injection because it instructs the agent to ingest and analyze potentially untrusted data from application interfaces, console logs, and external reports.

  • Socket1d

    No alerts

  • Snyktoday

    Risk: MEDIUM · 1 issue

Signed by skilld at b2e4283. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 12 hours ago.

Activeupdated yesterday

README badge

README badge for raphaelsalaja/skills/verify