agentleFS
DnsQueenDnsQueen/orcinus-skills/skills/independent-review/SKILL.md

independent-review

Skill3 starsChanged 7 days agoMITView on GitHub

Review a change independently before it merges, with six focused reviewers reading the diff — not the author's summary. Use before committing, before creating a pull request, and after a logical chunk of work. The implementer is never the final judge. Only invoke via the orcinus router.

14 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 10 edited versions we read keep these.

  • 2

    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:

    Only report an issue with confidence ≥ 80.

    line 20
  2. Distinctive:

    Every finding names a file and a line.

    line 21

“Common” is measured: kept by at least 6 in 10 of the 10 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

INDEPENDENT-REVIEW — the implementer is never the final judge

A change the author reviewed themself is not reviewed. This skill runs an independent review of a change: six focused reviewers read the diff and the code, each from one angle, and their findings are aggregated into one verdict. It does not fix code — fixing is a separate step, one slice at a time, test first.

Licences

This skill's text is MIT. The bundled reviewer agents and command (registered at the plugin root, agents/ and commands/) are Apache-2.0 — see ATTRIBUTION.md. Do not re-license the bundled files.

Rules

  • The reviewer reads the diff and the project's own conventions, never the author's summary. A summary that says "just a refactor" is a claim, not evidence.
  • Review only the diff. Never review the whole codebase — that hides the change in noise.
  • Only report an issue with confidence ≥ 80. A guess is not a finding.
  • Every finding names a file and a line. A finding without a location is not actionable.
  • A check that did not run is never reported as passed.

Steps, in order

1. Determine the scope

Pick the mode first, then list the changed files (new files included). That list is the scope. Nothing outside it is reviewed.

  • Local (uncommitted work): staged + unstaged + untracked files. git status alone misses staged changes, and git diff alone misses untracked files — list all three.
  • Single commit: the commit against its parent (git diff --name-only <commit>^ <commit>).
  • Pull request: the head against the merge-base of the target branch (git diff --name-only $(git merge-base <target> HEAD) HEAD). In a clean PR worktree without a base, comparing against the branch's merge-base is the only way to see the PR's changes — a plain git diff shows nothing.

2. Pick the applicable reviewers

  • Always: code-reviewer (general quality, project guidelines, bugs).
  • If test files changed, or if behaviour changed with no new tests: pr-test-analyzer (behavioral coverage, critical gaps, test quality). A behaviour change with no new test is itself a coverage finding — do not skip the analyzer because no test file changed.
  • If comments or docs were added: comment-analyzer (comment accuracy vs code, comment rot).
  • If error handling changed: silent-failure-hunter (silent failures, catch blocks, fallbacks).
  • If types were added or modified: type-design-analyzer (encapsulation, invariants).
  • After the review passes: code-simplifier (clarity, unnecessary complexity — preserves functionality).

3. Run the reviewers

The reviewers are the bundled agents registered at the plugin root (code-reviewer, code-simplifier, comment-analyzer, pr-test-analyzer, silent-failure-hunter, type-design-analyzer — from pr-review-toolkit, Apache-2.0). Run them in parallel when they are independent — they read the same diff from different angles. Each one reads the code, not a summary of it.

Confidence scales, per reviewer:

  • code-reviewer: 0–100; 91–100 is a critical bug or an explicit convention violation. Report only ≥ 80.
  • pr-test-analyzer: rates each gap 1–10 (10 = absolutely essential, must add).
  • silent-failure-hunter: CRITICAL (silent failure, broad catch), HIGH (poor error message, unjustified fallback), MEDIUM (missing context).
  • type-design-analyzer: rates four dimensions 1–10 — encapsulation, invariant expression, usefulness, enforcement.
  • comment-analyzer: flags accuracy and rot, with confidence.
  • code-simplifier: suggests simplifications; it never changes behaviour.

4. Aggregate

Collect the findings into one report:

  • Critical Issues (must fix before merge)
  • Important Issues (should fix)
  • Suggestions (nice to have)
  • Positive Observations (what is good)

A finding appears once, under its highest severity, with the reviewer that found it.

5. Action plan

Fix the critical issues first, then the important ones, then consider the suggestions. Re-run the review after the fixes — a fix that is not re-reviewed is not verified.

Report

Return the report in this form:

  • Scope (the changed files, from git diff).
  • Reviewers run (which ones, and why the others did not apply).
  • Critical Issues (every one, with file, line, reviewer, and confidence/severity).
  • Important Issues and Suggestions (every one, with file and line).
  • Positive Observations.
  • Verdict: PASS (no critical issues, and every reviewer that applied ran) or FAIL (a critical issue was found, or a reviewer that applied did not run).

Scope

Review only. This skill does not write code, does not fix findings, and does not decide the merge — a human does, after the critical issues are fixed and the review is re-run. The bundled agents carry project-specific references from the original author's project (a CLAUDE.md, Sentry/Statsig logging functions, an error-ids file). Read the project's own conventions (AGENTS.md / CLAUDE.md) instead of those hardcoded names; the method transfers, the names do not.

Sources

Method, in our own words, following pr-review-toolkit (anthropics/claude-plugins-official, Apache-2.0). The six reviewer agents and the review-pr command are bundled as-is at the plugin root (agents/, commands/); see ATTRIBUTION.md. No text was copied from the toolkit into this file.

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.