Home / All Skills / Workflow & Core Tools / caveman-review
English (EN)简体中文 (ZH)日本語 (JA)
Workflow & Core Tools · juliusbrussee/caveman

caveman-review

Compressed code review - one line per finding with location, problem and fix. Use for /caveman-review, "review this PR", or "review the diff".

L1 Basic · Source registration onlyV1 Commit-PinnedDual-pass Audit
Author juliusbrusseeGitHub juliusbrussee/caveman ↗Stars 104500License MIT + BSL-1.1 分拆（技能本体 MIT；GitHub API spdx NOASSERTION）。Commit 15581d1400
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Pure prompt-based review output compressor: Centered on an output format protocol—one line per finding: `L<line>: <problem>. <fix>.`, prefixed with `<file>:L<line>:` for multi-file diffs.

skills/caveman-review/SKILL.md
**Format:** `L<line>: <problem>. <fix>.` — or `<file>:L<line>: ...` when reviewing multi-file diffs.
Format protocol replaces free-form prose: The three-part structure of location + problem + fix compresses review comments into scannable single lines ready to paste directly into a PR.

Optional severity prefix levels: Four tiers—🔴 bug / 🟡 risk / 🔵 nit / ❓ q—allowing authors to filter by urgency.

skills/caveman-review/SKILL.md
`🔴 bug:` — broken behavior, will cause incident
Each of the four tiers specifies triggering criteria (causes an incident / fragile but functional / ignorable style nit / genuine question rather than suggestion); prefixes are used only when severity is mixed.

Anti-boilerplate phrasing rules: Explicitly lists fluff patterns to delete ('I noticed that…', 'You might want to consider…', repeating code lines, hedging) alongside corresponding ✅ replacements.

skills/caveman-review/SKILL.md
**Drop:** - "I noticed that...", "It seems like...", "You might want to consider..." - "This is just a suggestion but..." — use `nit:` instead -
Pairs each ❌ example with a ✅ single-line rewrite (e.g., 'I noticed that on line 42 you're not checking if the user object is null…' → 'L42: 🔴 bug: user can be null after .find(). Add guard before .email.'), training output habits through positive and negative contrast.

Preservation checklist ensures technical precision: Retains exact line numbers, backticks around symbol names, and concrete fixes instead of vague advice like 'consider refactoring'.

skills/caveman-review/SKILL.md
**Keep:** - Exact line numbers - Exact symbol/function/variable names in backticks - Concrete fix, not "consider refactoring this" -
Compresses tone and preamble without compromising locatability (line numbers/symbols) or actionability (specific fixes).

Auto-Clarity exceptions: Drops back to full paragraph prose for CVE-grade security findings, architectural disputes, or junior developer onboarding/mentoring, before returning to single-line mode.

skills/caveman-review/SKILL.md
Drop terse mode for: security findings (CVE-class bugs need full explanation + reference), architectural disagreements (need rationale, not just a one-liner), and onboarding contexts where the author is new and needs the "why".
Isomorphic with the Auto-Clarity mechanism in the primary caveman skill—high-risk scenarios or those requiring deeper context are not compressed.

Boundary discipline: Produces only review comment text—does not write code fixes, approve/reject PRs, or run linters; output is ready to paste.

skills/caveman-review/SKILL.md
Reviews only — does not write the code fix, does not approve/request-changes, does not run linters.
Positioned as a 'review expression style' rather than a review workflow: Operational actions (approve/request-changes/running tools) are explicitly out of scope; it serves strictly as output format guidance.

2Core Capabilities

01Single-line review comment output (`L<line>: <problem>. <fix>.`, with file prefix for multi-file diffs)
02Four-tier severity grading (bug/risk/nit/q)
03Elimination of review boilerplate and hedging (uses q: when uncertainty is necessary)
04Preservation of exact locations and concrete fix suggestions (line numbers, symbol names, specific solutions)
05Auto-Clarity: Automatically expands into full explanations for security, architecture, or teaching scenarios
06Ready-to-paste PR comments (does not approve PRs or execute tools)

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Can be auto-loaded
For Administrators
Register source and version
Evidence & Triggers (File Paths & Excerpts)
  • SKILL.md: Pure prompt instructions and rule text; no script execution or external impact.
Audit Notes: Pure prompt-based skill: zero scripts, zero network access, zero credentials, and zero local disk writes. Output consists solely of reformatted review text, and Auto-Clarity ensures critical security findings (CVEs) expand into full explanations rather than being forcibly compressed. Considerations are limited to conversational guidance with no code-level risks.

5Second-Pass Verification

  • [ok] 纯 prompt、无脚本 — skills/caveman-review 目录 git ls-tree 仅 SKILL.md + README.md；无任何可执行文件。
  • [ok] 单行格式协议 — 原文 '**Format:** `L<line>: <problem>. <fix>.`' 存在；多文件变体 `<file>:L<line>:` 亦在。
  • [ok] 四档严重度 — 🔴 bug / 🟡 risk / 🔵 nit / ❓ q 四行定义逐字存在。
  • [ok] Drop/Keep 清单 — 'Drop:' 下列出 I noticed that… 等套话与复述代码行；'Keep:' 下列出精确行号/符号名/具体修法。
  • [ok] Auto-Clarity 例外 — 原文含 security findings (CVE-class…)/architectural disagreements/onboarding contexts 三类展开条件。
  • [ok] 边界声明与源码一致 — 'does not write the code fix, does not approve/request-changes, does not run linters' 逐字在——无任何工具调用指令。
  • [ok] 外部依赖为空 — 文件级扫描 URL/CLI/API 零命中；无 linter/CI/评审平台依赖。
  • [ok] meta 元数据 — 与 caveman 同仓库：stars 104500、pushed_at 2026-09-09、本地 HEAD==origin/main==pin 15581d1；技能 MIT 面。

6Verdict & Scenarios

  • Highly actionable output protocol: Single-line 'location/problem/fix' makes review comments scannable, filterable, and directly pasteable.
  • Severity grading minimizes noise: Authors can focus exclusively on 🔴 bug and ignore 🔵 nit.
  • Concrete anti-boilerplate rules with paired examples make model alignment easy.
  • Zero compression for security findings: CVE-level items receive full explanations and references.
  • Clean boundary of responsibility: Strictly an output style guide without overstepping into code editing, approvals, or tool execution.
    • Best for: Ideal for developers seeking a 'short, scannable, directly pasteable into PRs' review style during code review sessions (/caveman-review, 'review this PR', 'review the diff'); well-suited for reviewers and authors already familiar with the codebase who only need location and fix pointers.
    Not for: Not suitable for security audits or architectural reviews requiring comprehensive justification (Auto-Clarity expands specific items, but overall output remains single-line oriented); not suitable for pedagogical reviews aimed at beginners (expanding every item defeats the purpose during onboarding); not suitable for review workflows expecting automated linter execution, PR approvals, or code commits—it explicitly outputs text only.
    Install Agent-Ready
    ① Verified Mirror 2026-09-09
    Option A · Direct DownloadDownload caveman-review.tar.gz
    sha256: 94f8129bf1ddbecf…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit 15581d1400.