Name: claude-review
Author: yangliu2060

Imported from yangliu2060/smith--skills (claude-review/SKILL.md). Install upstream with npx skills add yangliu2060/smith--skills --skill claude-review. Copyright stays with the author.

claude-review — Claude 无界面评审

用一次 claude --print 获得 Claude 的独立评审，不开启完整交互会话。

When To Invoke

Situation	Type
Detailed implementation plan, about to start coding	`plan`
Module boundaries / system diagram drafted	`architecture`
README / API docs / user-facing docs before shipping	`docs`
STATUS.md / errors.md after a work session — verify honesty	`work-log`

Usage

~/.codex/skills/claude-review/scripts/claude-review <type> <file_or_dir> [--deep]

Examples:

# Review plan before implementation (auto-chains plan-eng-review)
~/.codex/skills/claude-review/scripts/claude-review plan ./docs/plan.md

# Architecture review with deep consultation (also chains office-hours)
~/.codex/skills/claude-review/scripts/claude-review architecture ./docs/design.md --deep

# Quality-check docs before shipping (single pass, no chain)
~/.codex/skills/claude-review/scripts/claude-review docs ./README.md

# Audit a work directory for buried failures
~/.codex/skills/claude-review/scripts/claude-review work-log ./.codexfree/

What Happens

Script validates inputs and resolves absolute path
Assembles a headless-mode prompt (no AskUserQuestion, single final report)
Calls claude --print --permission-mode bypassPermissions once
For plan / architecture: Claude auto-invokes its plan-eng-review skill in the same session and consolidates
For --deep (plan / architecture only): Claude additionally chains the office-hours skill
Report streamed to stdout AND saved to .claude-review/<timestamp>-<type>.md in the current working directory

All skill chaining happens inside ONE Claude session — you pay for one --print invocation regardless of how many sub-skills are chained.

Output Format

The report is always Markdown with this structure:

# Review Report — <type>
**Target:** <absolute path>
**Timestamp:** <UTC ISO>

## Executive Summary
...

## Critical Issues (must fix before proceeding)
...

## High-Priority Findings (should fix soon)
...

## Suggestions (nice to have)
...

## Chained: plan-eng-review  (only for plan/architecture)
...

## Chained: office-hours  (only with --deep)
...

## Decision
approve | revise | block
<1-2 sentence rationale>

Codex should read the report, treat revise findings as required fixes, then re-invoke the skill on the updated artifact. block means the artifact has fundamental problems and needs a rewrite, not a patch.

Decision Taxonomy

Decision	Meaning	Codex action
`approve`	Artifact is ready to proceed	Continue to next phase
`revise`	Issues found, fixable in place	Fix issues, re-invoke skill
`block`	Fundamental problem, not patchable	Halt; escalate to user

Why Headless?

Calling Claude interactively burns tokens on greetings, clarification, and tool overhead. --print mode:

Single request → single response → exit
All tool use happens inside one session
No context carried across invocations
Much lower token cost than a conversation

The tradeoff: Claude cannot ask follow-up questions. The script's prompt assumes sensible defaults and instructs Claude to produce a complete report in one pass.

Skill Chain Reference

Type	Initial review	Chains `plan-eng-review`	Chains `office-hours`
`plan`	✅	✅	only with `--deep`
`architecture`	✅	✅	only with `--deep`
`docs`	✅	❌	❌
`work-log`	✅	❌	❌

docs and work-log don't chain because plan-eng-review is an architecture-review skill, not a copy-editing or audit skill.

Prerequisites

claude CLI installed (npm install -g @anthropic-ai/claude-code)
User is logged into Claude Code (claude --version works)
For plan / architecture: the plan-eng-review skill must be installed in the user's Claude skill directory (part of the gstack plugin)
For --deep: the office-hours skill must be installed

Files

scripts/claude-review — main bash wrapper
references/review-prompts.md — lens definitions for each review type (Claude reads these via the bash-assembled prompt)

Install claude-review and it becomes one more candidate the router can pick - when it fits.

A listing is a routing target with a manifest: what it does, which domains it covers, what it costs and who publishes it. Once installed it sits beside your own models and tools, is scored like any other candidate for each request, and shows up in the trace when it wins. Ratings come from workspaces that installed it, one per account.

{ "ocm": "1", "id": "yangliu2060-smith-skills-claude-review", "kind": "skill", "name": "claude-review", "description": "Claude 无界面评审：用一次 `claude --print` 审查 plan、架构文档、用户文档或 work-log，并返回结构化报告。", "publisher": "yangliu2060", "version": "1.0.0", "capabilities": { "domains": [ "general" ], "tags": [ "skill-md", "skills-sh" ], "languages": [ "en" ] }, "quality_prior": 0.6, "examples": [ "Claude 无界面评审：用一次 `claude --print` 审查 plan、架构文档、用户文档或 work-log，并返回结构化报告。" ], "primary": false, "metadata": { "source": { "provider": "skills.sh", "repository": "https://github.com/yangliu2060/smith--skills", "path": "claude-review/SKILL.md", "ref": "HEAD", "url": "https://github.com/yangliu2060/smith--skills/blob/HEAD/claude-review/SKILL.md", "key": "yangliu2060/smith--skills/claude-review/SKILL.md" } }, "instructions": "# claude-review — Claude 无界面评审\n\n用一次 `claude --print` 获得 Claude 的独立评审，不开启完整交互会话。\n\n## When To Invoke\n\n| Situation | Type |\n|---|---|\n| Detailed implementation plan, about to start coding | `plan` |\n| Module boundaries / system diagram drafted | `architecture` |\n| README / API docs / user-facing docs before shipping | `docs` |\n| STATUS.md / errors.md after a work session — verify honesty | `work-log` |\n\n## Usage\n\n```bash\n~/.codex/skills/claude-review/scripts/claude-review <type> <file_or_dir> [--deep]\n```\n\n**Examples:**\n\n```bash\n# Review plan before implementation (auto-chains plan-eng-review)\n~/", "cost": { "context_tokens": 1037 } }

claude-review

About