eranw2000/claude-review-agents

Claude Code review agents

Four Claude Code agents that check your work before it ships: one for code quality, one for writing, one for release-readiness across common stacks, and one that runs a pull request's tests and certifies whether it is ready to release. A companion hook runs a quick version of the writing check on every file Claude writes.

The flow

Source: docs/review-agents-flow.drawio (editable in draw.io).

How these work

A Claude Code agent is a subagent with its own context window and a restricted set of tools. Claude can invoke one on its own when the description matches what you are doing, or you can ask for it by name. Each runs its review and reports back without editing your files, so the noise of a review stays out of your main conversation and you keep control of the changes.

• code-reviewer: senior code review for quality and maintainability. Reads the pending git diff, reports findings as Critical / Warning / Suggestion with file:line and a concrete fix. Looks at clarity, duplication, error handling, edge cases, over-large functions, test coverage, and consistency with the surrounding style. For changes to parsers, gates, and dispatchers it reproduces findings by executing the real path instead of reasoning about the code, and it checks new tests for vacuity (tests that cannot fail). Reviews only; does not edit.
• ai-signal-reviewer: checks human-facing prose (docs, READMEs, PR descriptions, commit messages, emails) for AI writing tells and formatting traps: em dashes, AI-typical vocabulary, significance inflation, rule-of-three, elegant variation, curly quotes, and Unicode box-drawing characters in tables. Flags each issue with its location and a short replacement rather than rewriting the piece. Runs on the sonnet model (the routine-work tier); change the model: line in the file if you want a different one.
• deploy-guard