Claude Cursor Skill

reviewing-code

Imported from alexei-led/cc-thingz/src/skills/reviewing-code.

LLM Mart · 0 points · 2 views 0 listing impressions 0 install-command copies
Download alexei-led-cc-thingz-src_skills_reviewing-code-0202df8.zip · 9 KB

Install

skills CLI npx skills add https://github.com/alexei-led/cc-thingz/tree/master/src/skills/reviewing-code
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install alexei-led-cc-thingz@llmmart
Git git clone https://github.com/alexei-led/cc-thingz.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole alexei-led/cc-thingz collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Code Review

Produce findings, not edits, for the requested diff, PR, changed files, or file list. Read references/severity-rubric.md before scoring or reporting: it owns the dimensions, severity, confidence, and score rules.

Load a language reference only when a finding needs language or framework semantics; do not read every reference merely because that language is in scope:

  • C#: references/csharp.md
  • Go: references/go.md
  • Java/Kotlin: references/java-kotlin.md
  • Python: references/python.md
  • Rust: references/rust.md
  • TypeScript: references/typescript.md
  • Web, HTML, CSS, JS, HTMX: references/web.md

Other languages: use the rubric alone; report reduced coverage if missing language guidance limits the review.

Workflow ownership

This skill owns the finding rubric and domain references. Coding-posture owns the general review checklist. Use Revmux for explicitly requested multi-agent review when installed, and Revdiff for explicitly requested interactive annotations; neither is required for an ordinary review. Do not run several review workflows for one request unless the user asks.

Scope

Use the scope the user named. Otherwise ask one question offering:

  • uncommitted changes
  • branch compared to the default branch
  • specific files or a PR diff

Without command access, work from the supplied diff and files, and ask for them if absent. No changes in scope: report Nothing to review.

Modes

Default is standard.

  • Quick: small, low-risk diffs. Security and correctness; changed lines plus direct context.
  • Standard: security, correctness, and tests; follow callers or callees when a finding needs them.
  • Deep: on request for a thorough, risk, or merge-safety review. All rubric dimensions, including boundary inputs and affected tests.
  • Team: on request for team, parallel, or multi-pass review, or offered for a large diff. Split by dimension or file group, then merge into one report: dedupe by file:line plus claim, keep the higher severity only when evidence supports it, and put disagreements under Needs review.
  • External: only when the user explicitly asks for an external, second-model, or second-opinion review. Keep private code local unless the bridge runs locally or the user approved sharing. Rate external output with the same rubric; claims you cannot verify go to Needs review. Report whether it completed, was unavailable, or was skipped and why.
  • Simplify: on request for an over-engineering or "what can we delete" review. Simplicity dimension only. Instead of the Output template, one line per finding: file:line — <tag> <what>. <replacement>., with tags delete, stdlib/native (name the replacement), yagni, shrink. End with net: -N lines possible. or Lean already. Ship.

Evidence

  • Prove each claim from source or observed checks. A read-only reviewer reports checks to run; the caller or engineer executes them. Do not present a proposed check as a passing result.
  • If a code-graph tool (GitNexus, codegraph) is installed, use it for caller, impact, and test-coverage questions on broad or public-API changes; a stale index is a coverage gap, not evidence.
  • Web research is for public facts only (CVEs, official docs). Never send private code or diffs to web tools.
  • Skip style that project tooling already enforces. List suspicious code outside scope as out of scope instead of widening the review.
  • Review generated or vendored files only when source generation is in scope.

Output

## Code Review Summary

Scope: <description>
Depth: quick | standard | deep | team | external
Languages: <list>
Coverage: complete | partial — <reason>
Graph evidence: none | <tool> — <freshness/gaps>
External review: not requested | completed | unavailable | skipped — <reason>
Score: <N/10 if requested> — confidence <high|medium|low>

### Critical

- `file:line` — <category>, confidence <level>. <issue> Scenario: <how it fails>. Fix: <concrete fix>.

### Warnings

- `file:line` — <category>, confidence <level>. <issue> Scenario: <how it fails>. Fix: <concrete fix>.

### Suggestions

- `file:line` — <category>, confidence <level>. <improvement>. Fix: <concrete fix>.

### Needs review

- `file:line or tool/context gap` — <missing context and why it matters>.

### Summary

<2-3 sentences: merge risk and next actions, or "No confirmed findings".>

Omit empty sections, except Needs review when it explains partial coverage.

Platform additions

No target-specific additions.

Files (cc-thingz)
  • .agentbundler
    • targets
      • claude.json 1.3 KB
        Loading…
  • references
    • csharp.md 772 B
      Loading…
    • go.md 820 B
      Loading…
    • java-kotlin.md 1 KB
      Loading…
    • python.md 784 B
      Loading…
    • rust.md 1 KB
      Loading…
    • severity-rubric.md 4.5 KB
      Loading…
    • typescript.md 1007 B
      Loading…
    • web.md 782 B
      Loading…
  • SKILL.md 5 KB
    ---
    description:
      Use when reviewing changed code, PRs, diffs, or specific files. Finds
      evidence-backed defects in security, correctness, tests, reliability, performance,
      maintainability, and docs. Supports quick, standard, deep, team, and external-review
      modes, plus a simplify mode for over-engineering and "what can we delete" reviews. NOT for repo-wide architecture review, general codebase exploration, fixing
      issues (use fixing-code), or improving tests without a code review (use improving-tests).
    name: reviewing-code
    ---
    
    # Code Review
    
    Produce findings, not edits, for the requested diff, PR, changed files, or file
    list. Read `references/severity-rubric.md` before scoring or reporting: it owns
    the dimensions, severity, confidence, and score rules.
    
    Load a language reference only when a finding needs language or framework
    semantics; do not read every reference merely because that language is in scope:
    
    - C#: `references/csharp.md`
    - Go: `references/go.md`
    - Java/Kotlin: `references/java-kotlin.md`
    - Python: `references/python.md`
    - Rust: `references/rust.md`
    - TypeScript: `references/typescript.md`
    - Web, HTML, CSS, JS, HTMX: `references/web.md`
    
    Other languages: use the rubric alone; report reduced coverage if missing
    language guidance limits the review.
    
    ## Workflow ownership
    
    This skill owns the finding rubric and domain references. Coding-posture owns the
    general review checklist. Use Revmux for explicitly requested multi-agent review
    when installed, and Revdiff for explicitly requested interactive annotations;
    neither is required for an ordinary review. Do not run several review workflows
    for one request unless the user asks.
    
    ## Scope
    
    Use the scope the user named. Otherwise ask one question offering:
    
    - uncommitted changes
    - branch compared to the default branch
    - specific files or a PR diff
    
    Without command access, work from the supplied diff and files, and ask for
    them if absent. No changes in scope: report `Nothing to review.`
    
    ## Modes
    
    Default is standard.
    
    - **Quick**: small, low-risk diffs. Security and correctness; changed lines plus direct context.
    - **Standard**: security, correctness, and tests; follow callers or callees when a finding needs them.
    - **Deep**: on request for a thorough, risk, or merge-safety review. All rubric dimensions, including boundary inputs and affected tests.
    - **Team**: on request for team, parallel, or multi-pass review, or offered for a large diff. Split by dimension or file group, then merge into one report: dedupe by `file:line` plus claim, keep the higher severity only when evidence supports it, and put disagreements under Needs review.
    - **External**: only when the user explicitly asks for an external, second-model, or second-opinion review. Keep private code local unless the bridge runs locally or the user approved sharing. Rate external output with the same rubric; claims you cannot verify go to Needs review. Report whether it completed, was unavailable, or was skipped and why.
    - **Simplify**: on request for an over-engineering or "what can we delete" review. Simplicity dimension only. Instead of the Output template, one line per finding: `file:line — <tag> <what>. <replacement>.`, with tags `delete`, `stdlib`/`native` (name the replacement), `yagni`, `shrink`. End with `net: -N lines possible.` or `Lean already. Ship.`
    
    ## Evidence
    
    - Prove each claim from source or observed checks. A read-only reviewer reports checks to run; the caller or engineer executes them. Do not present a proposed check as a passing result.
    - If a code-graph tool (GitNexus, codegraph) is installed, use it for caller, impact, and test-coverage questions on broad or public-API changes; a stale index is a coverage gap, not evidence.
    - Web research is for public facts only (CVEs, official docs). Never send private code or diffs to web tools.
    - Skip style that project tooling already enforces. List suspicious code outside scope as out of scope instead of widening the review.
    - Review generated or vendored files only when source generation is in scope.
    
    ## Output
    
    ```markdown
    ## Code Review Summary
    
    Scope: <description>
    Depth: quick | standard | deep | team | external
    Languages: <list>
    Coverage: complete | partial — <reason>
    Graph evidence: none | <tool> — <freshness/gaps>
    External review: not requested | completed | unavailable | skipped — <reason>
    Score: <N/10 if requested> — confidence <high|medium|low>
    
    ### Critical
    
    - `file:line` — <category>, confidence <level>. <issue> Scenario: <how it fails>. Fix: <concrete fix>.
    
    ### Warnings
    
    - `file:line` — <category>, confidence <level>. <issue> Scenario: <how it fails>. Fix: <concrete fix>.
    
    ### Suggestions
    
    - `file:line` — <category>, confidence <level>. <improvement>. Fix: <concrete fix>.
    
    ### Needs review
    
    - `file:line or tool/context gap` — <missing context and why it matters>.
    
    ### Summary
    
    <2-3 sentences: merge risk and next actions, or "No confirmed findings".>
    ```
    
    Omit empty sections, except Needs review when it explains partial coverage.
    
    ## Platform additions
    
    No target-specific additions.
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.