Refute
The author of a change shares its assumptions, so a review by the same context tends to approve it. Review from a fresh context, against the goal, and try to refute every finding, and every claim of success, before it counts.
A review can end clean. If there is no real defect, the answer is no defects found, plus what you checked. Do not manufacture findings to look useful.
Who reviews
- Default:
dojo-protocol:reviewer (sonnet). Fresh context, Read/Glob/Grep/Bash. Give it the goal and the files, not your conclusion.
- Claims that need a hard second look (an agent's "all green", a risky merge):
dojo-protocol:judge (opus). It tries to refute, and it re-runs checks itself.
- The reviewer is read-only by instruction. Its tools list is the only hard limit; Bash can still write if told to.
- When a judge or reviewer checks work against rules, paste the rules verbatim with their source path and line, never a paraphrase.
Steps
- Scope. Collect the diff or the files, and read the goal. If there is nothing to review, say so and stop.
- Read in full. Read whole files, not only the changed lines; bugs sit where new code meets old.
- Pass with three lenses, in one pass, as prompts and not as a quota:
- Breaks in production: worst input, failing dependency, double run, concurrent run, swallowed error, resource left open.
- Hard to maintain: names that hide intent, hidden coupling, magic values, tests that pin implementation instead of behavior.
- Security: untrusted input reaching a command or query, secrets in code or logs, missing access checks, new attack surface.
- Refute each finding. A finding stays only if it has a
path:line and a check: a command, a failing test, or a concrete input that shows the problem. Run the check. If it cannot be shown, drop it from the findings and log it as a suspicion with the reason.
- Rank what survives: critical (data loss, security hole, broken build), high, medium, low.
- Verdict. Either
no defects found (with what you checked) or the ranked list.
Output
Verdict: no defects found | defects found
Checked: <files read, commands run, results>
Findings:
1. [high] path/file.py:42 - what breaks - how to reproduce
Dropped suspicions: <what, and why it could not be shown>
Returned: <reviewers dispatched / returned>
Traps
- A dead reviewer looks the same as a clean one. Count returns: if a reviewer produced nothing, it did not review.
- An agent's verdict is a claim. When it says clean, ask what it ran.
- A check built from the code under test cannot fail where that code is wrong. Prefer an independent input.
- Gates raise the floor; they do not certify the result. Say what the review did not cover.
- Review the goal as well as the code: a change can be correct and still solve the wrong problem.
Related: rule 5 and rule 6 in ${CLAUDE_PLUGIN_ROOT}/PROTOCOL.md; delegate for briefing the reviewer.
1---2name: refute3description: "Use when reviewing a change or testing a claim: a finding needs path:line and a check that shows it; a clean pass says so."4model: inherit5category: repo-docs-health6metadata:7 source: https://skillmd.com/skills/dojogenesis/refute8---910# Refute1112The author of a change shares its assumptions, so a review by the same context tends to approve it. Review from a fresh context, against the goal, and try to refute every finding, and every claim of success, before it counts.1314A review can end clean. If there is no real defect, the answer is `no defects found`, plus what you checked. Do not manufacture findings to look useful.1516## Who reviews1718- Default: `dojo-protocol:reviewer` (sonnet). Fresh context, Read/Glob/Grep/Bash. Give it the goal and the files, not your conclusion.19- Claims that need a hard second look (an agent's "all green", a risky merge): `dojo-protocol:judge` (opus). It tries to refute, and it re-runs checks itself.20- The reviewer is read-only by instruction. Its tools list is the only hard limit; Bash can still write if told to.21- When a judge or reviewer checks work against rules, paste the rules verbatim with their source path and line, never a paraphrase.2223## Steps24251. **Scope.** Collect the diff or the files, and read the goal. If there is nothing to review, say so and stop.262. **Read in full.** Read whole files, not only the changed lines; bugs sit where new code meets old.273. **Pass with three lenses**, in one pass, as prompts and not as a quota:28 - *Breaks in production:* worst input, failing dependency, double run, concurrent run, swallowed error, resource left open.29 - *Hard to maintain:* names that hide intent, hidden coupling, magic values, tests that pin implementation instead of behavior.30 - *Security:* untrusted input reaching a command or query, secrets in code or logs, missing access checks, new attack surface.314. **Refute each finding.** A finding stays only if it has a `path:line` and a check: a command, a failing test, or a concrete input that shows the problem. Run the check. If it cannot be shown, drop it from the findings and log it as a suspicion with the reason.325. **Rank** what survives: critical (data loss, security hole, broken build), high, medium, low.336. **Verdict.** Either `no defects found` (with what you checked) or the ranked list.3435## Output3637```38Verdict: no defects found | defects found39Checked: <files read, commands run, results>40Findings:41 1. [high] path/file.py:42 - what breaks - how to reproduce42Dropped suspicions: <what, and why it could not be shown>43Returned: <reviewers dispatched / returned>44```4546## Traps4748- A dead reviewer looks the same as a clean one. Count returns: if a reviewer produced nothing, it did not review.49- An agent's verdict is a claim. When it says clean, ask what it ran.50- A check built from the code under test cannot fail where that code is wrong. Prefer an independent input.51- Gates raise the floor; they do not certify the result. Say what the review did not cover.52- Review the goal as well as the code: a change can be correct and still solve the wrong problem.5354Related: rule 5 and rule 6 in `${CLAUDE_PLUGIN_ROOT}/PROTOCOL.md`; `delegate` for briefing the reviewer.