independent-review makes sure the session that wrote the code never clears it. A wrapper script starts a fresh Codex, headless Claude Code or Grok reviewer with read-only tools and no connectors or memory, feeds it a subject computed from git, and refuses a short round that only looks clean.
The rest of the skill is method: how to write a brief that does not steer the reviewer, how to triage each finding by what goes wrong for a person and how often, and when to stop.
What You Get
- A wrapper command, review.py, that picks the reviewer, applies read-only flags and fails a round that is too short or lacks read evidence
- A subject definition computed from git, with rules for splitting large ranges by commits or by surface
- A five-part brief template and a triage procedure covering severity, frequency and the lines a finding must cross
- A stopping rule: no introduced finding above Low and no High or Critical of any age
- A reference on what each reviewer CLI still loads and how to run a round by hand
Common Use Cases
- Reviewing a phase of AI-built work before pushing to a deploying branch
- A whole-history review before a first release
- Re-reviewing after each fix until a round is clean
- Deciding which reviewer findings are real, rare or false
What Makes It Unique
It treats reviewer output and severity labels as claims to verify, and it builds the failure it found into the tooling: a round that exits 0 after one sentence is rejected. It requires a fresh session of the reviewer’s own, with fallback order across three model families, and records what could not be switched off. Its triage favors removing a rare case over adding machinery and records accepted Lows.
Under The Hood
The skill is a long markdown file, two reference files and one Python script. The script runs git commands to build the subject, writes the subject and prompt into the git directory, and launches a Codex, Claude Code or Grok command line in read-only mode with tools limited to reading and searching. It checks the reviewer’s event stream, reads binary location environment variables, and saves the report and events file where the user points. It makes no web calls of its own, and the reviewer’s web search is switched off. The agent also runs git status, log and diff itself.