Witness proof

Skill bendaamerahmed/witness/skills/witness-proof

Produce the evidence block for a change you just made, or are about to claim is done. Walks the chain of custody out loud, reproduce, locate, repair, re-run, widen, and reports honestly where it stopped. Use when the user asks for proof, receipts, or verification that something actually works, or before reporting a bug fix. Trigger: /witness-proof, "witness proof", "prove it", "show me the receipts", "did you actually run it", "verify this fix", "how do you know it works".From its SKILL.md

Install

npx -y skills add bendaamerahmed/witness --skill witness-proof

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its file declares

Copied from the file, not written here

The file declares its own license as Apache-2.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

2.6 KB, as published. Nobody here has run it

witness-proof

Produce the evidence for one change. Not a report about the change, the evidence.

Procedure

Reproduce. Identify the smallest command that demonstrates the problem. Run it. Quote the real output. If the change is a feature rather than a fix, the reproduction is the check that fails because the feature does not exist yet.
Locate. Name the file and line that produces the wrong behavior. One sentence on why it is wrong.
Repair. Point at the diff. Confirm out loud that no check was modified. If one was, that is the headline, not a footnote, see the escalation below.
Re-run. The same command from step 1, unchanged. Quote the real output.
Widen. One input the original check did not cover, in the same shape as the bug. Quote the result.

Output

repro:  <command>  ->  <the failure, quoted>
cause:  <file:line>  <one sentence>
fix:    <file:line>  checks unchanged
proof:  <same command>  ->  <the pass, quoted>
widen:  <the extra case>  ->  <result>

Every line is either real output or the sentence "I did not run this". Never a paraphrase, never a prediction, never a description of what the output would look like.

When you cannot run it

Say it in the first line, not the last. Then produce the block anyway with the commands filled in and the outputs marked not run. A block full of not run is a useful artifact, it tells the reader exactly what to execute. A block full of invented outputs is worse than no block.

Escalation

If step 3 finds that a check was modified, stop the block and lead with this instead:

CHECK CHANGED — <file:line>
was:    <the old assertion>
now:    <the new assertion>
why:    <what was wrong with the old one>
proof:  the new check fails against the pre-fix code  ->  <quoted output>

A new check that never failed against the bug proves nothing. If you cannot produce that last line, the check change is not justified yet.

Boundaries

This skill does not review style, architecture or performance. It answers exactly one question: is there evidence for the claim.