review-code-with-evidence
Skill1 starsChanged 2 months agoMITView on GitHub
Read-only review of a PR, commit, diff, patch, config, or migration for reachable, evidenced defects with severity and one verdict. Not for diagnosing failures, writing fixes, or the push or merge itself.
In this repository21 other files it gives its agents.
- change-production-data-safelySkill
- craft-contextual-uiSkill
- diagnose-root-causesSkill
- engineer-code-changeSkill
- engineer-reliable-data-pipelinesSkill
- engineer-secure-systemsSkill
- evaluate-agent-behaviorSkill
- handle-sensitive-data-safelySkill
- infer-intent-before-codingSkill
- migrate-contracts-safelySkill
- operate-production-safelySkill
- orchestrate-long-running-workSkill
- plan-decision-complete-changeSkill
- ponytailSkill
- publish-verified-github-changeSkill
- release-production-changesSkill
- respond-to-production-incidentsSkill
- route-ultracode-subagentsSkill
- verify-end-to-endSkill
- write-durable-documentationSkill
- write-tests-that-prove-behaviorSkill
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
- Distinctive: line 12
Read repository and review instructions.
- Distinctive: line 14
Read the complete diff first, then the callers, tests, types, configuration, and contracts needed to understand runtime behavior.
- Distinctive: line 51
APPROVED: no validated finding or unmet condition remains, coverage is complete, and no material uncertainty could change the gate decision.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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.
Files
Review Code With Evidence
Find reachable defects and unmet requirements in a defined change set. Be skeptical of both the author's claims and suspicious-looking code; verify before reporting. A review request is read-only: do not edit files, resolve threads, post reviews, run destructive checks, or mutate external state.
Freeze the boundary
- Read repository and review instructions.
- Resolve the exact candidate, base, scope, intended outcome, and acceptance conditions from repository evidence, not guesses.
- Read the complete diff first, then the callers, tests, types, configuration, and contracts needed to understand runtime behavior.
- Separate behavior introduced or worsened by the candidate from pre-existing issues. Report pre-existing defects only when the scope includes them or they explain an introduced risk.
When the review is an independent gate for another agent's candidate, pin an immutable candidate before reading (commit and tree for Git; complete digest, baseline, and file inventory for a patch), take the sanitized task contract and raw diff but not the author's conclusions or desired verdict, and return BLOCKED rather than inventing identifiers. Follow the independent-gate procedure. The same model counts as independent only when it did not implement the change and starts without inherited implementer context; that is agent evidence, never human or GitHub approval.
Search by failure class
In priority order:
- incorrect behavior and missed acceptance conditions;
- data loss, corruption, races, stale state, transaction errors;
- authentication, authorization, privacy, injection, secret exposure, trust-boundary failures;
- public contract, schema, compatibility, migration, rollout, and rollback failures;
- error, retry, timeout, cancellation, cleanup, and partial-success paths;
- plausible hot-path or scale regressions;
- frontend accessibility, responsive, and state-handling regressions;
- complexity only when it creates a concrete defect risk.
Ignore taste, formatting, and speculative nits unless they cause a material problem or violate an explicit rule. Missing tests alone are not a defect; report the unprotected behavior only when it creates concrete regression risk.
Prove each finding
For every candidate finding: identify the tightest changed location; construct a reachable input, state, timing, or call path; inspect the guards, callers, configuration, and tests that might prevent it; compare expected with actual behavior, running a safe focused check when useful; establish impact and whether the candidate introduced or worsened it. Drop findings the evidence contradicts.
Report
Lead with validated findings sorted by severity:
- P0 Critical: catastrophic, widespread, or irreversible harm.
- P1 High: serious security, data, availability, or core-function failure that should block release.
- P2 Medium: realistic incorrect behavior or meaningful regression that should be fixed.
- P3 Low: bounded but real harm; never optional polish.
Each finding: specific title, file and line, concrete trigger and result, violated contract, evidence, smallest remediation direction. One defect per finding. If nothing remains, say so plainly and list only material residual risks or verification gaps.
For an independent gate, end with a coverage ledger (every assigned file, acceptance condition, material risk, and verification command) and exactly one verdict bound to the immutable candidate, and to the operation descriptor when the review gates a publication:
- APPROVED: no validated finding or unmet condition remains, coverage is complete, and no material uncertainty could change the gate decision.
- CHANGES_REQUIRED: at least one validated finding or unmet condition must be resolved.
- BLOCKED: candidate identity, scope, context, reviewer competence, or material evidence is insufficient to decide.
Automated checks alone never justify APPROVED. Partial scope never approves the whole candidate; hand security, privacy, billing, data, concurrency, or accessibility risk outside your competence to a qualified specialist review.
Stop and ask when
- the candidate, base, or scope cannot be pinned from evidence;
- a material acceptance or authorization decision is hidden behind redaction;
- acting on a finding is requested inside a review-only task (it needs separate authority).
Done when
Findings are evidenced and ranked, coverage is stated, and the verdict is understood to go stale on any change to the reviewed identity, ancestry, tree, delivered byte, dependency, generated artifact, or operation descriptor.
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
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.

