agentleFS
darrenCWJdarrenCWJ/agent-skills-scaffold/.claude/skills/verification/SKILL.md

verification

Skill0 starsChanged 3 months agoMITView on GitHub

Prove that a claim, output, or fix actually works using evidence, not confidence. Use whenever about to say "done", "fixed", "works", "passes", or "should be fine" — any completion claim needs this skill first, even for small changes and even when the work obviously looks right.

47 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use evidence-driven-writing, experiment-results-planning and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • Norman-burySkill

    evidence-driven-writing

    Use when writing or revising Introduction, Related Work, background, literature synthesis, or any section where references must drive claims

    Also in 3 projects that use this

  • Norman-burySkill

    experiment-results-planning

    Use when designing experiments, result tables, mock planning data, evaluation protocols, or results sections before real data are final

    Also in 3 projects that use this

  • +10 more from Norman-bury

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 5 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 39 edited versions we read keep these.

  • 5

    Distinctive

    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:

    Verify the requirement, not the implementation: the question is "does it do what was asked", not "does my code run".

    line 23
  2. Distinctive:

    Test the failure path too: a verifier that passes everything is broken.

    line 24
  3. Distinctive:

    Verify at boundaries: empty input, maximum size, wrong type, missing dependency, non-ASCII text, concurrent/repeated invocation — whichever …

    line 25
  4. Distinctive:

    After ANY later change (cleanup, formatting, "one tiny fix"), re-verify.

    line 26
  5. Distinctive:

    Every factual claim gets a source or a computation behind it.

    line 30

“Common” is measured: kept by at least 6 in 10 of the 39 edited versions of this file we could read. “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

Verification

Confidence is not evidence. A thing is verified when you have run a check that COULD have failed and watched it pass. Verification that cannot fail is theater.

Pick the strongest available verifier

Strongest first — drop down only when the stronger option genuinely doesn't exist:

  1. Execution — run the code, command, or script; observe actual output. For files: open/parse the real artifact (render the docx, run the notebook, load the CSV).
  2. Automated checks — tests, linters, type checkers, schema validators, link checkers. Run from clean state; cached green is not green.
  3. Independent recomputation — redo the calculation a different way (different tool, different method) and compare. Never verify math by re-reading the math; compute it programmatically.
  4. Fresh-agent inspection — subagent with no authorship bias checks the deliverable against the requirement.
  5. Line-by-line self-review — weakest; acceptable only for trivial changes, and still against the original requirement, not memory of it.

Rules of evidence

  • Record WHAT was run, the ACTUAL result, and WHEN. "Tests pass" without the command and output is a claim, not evidence.
  • Verify the requirement, not the implementation: the question is "does it do what was asked", not "does my code run".
  • Test the failure path too: a verifier that passes everything is broken. When feasible, break the thing intentionally once to confirm the check catches it.
  • Verify at boundaries: empty input, maximum size, wrong type, missing dependency, non-ASCII text, concurrent/repeated invocation — whichever apply.
  • After ANY later change (cleanup, formatting, "one tiny fix"), re-verify. The most embarrassing failures ship in the commit after the verified one.

Verifying claims in content (research, docs, analysis)

  • Every factual claim gets a source or a computation behind it. Mark anything unverifiable as an assumption, visibly.
  • Numbers: recompute from raw data. Quotes: check against the original. Links: fetch them.
  • Dates and "current" facts: check against today's date, not training memory.

Output

A verification record (in the plan's ledger, the critique file, or the final-check report): check performed → command/method → actual result → verdict.

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.