Table of Contents
One command. Three legendary perspectives. Zero mercy.
Powered by Claude Code Skills.
See it in action →
Features
- Four slash commands —
/elon-review, /jobs-review, /linus-review for solo perspectives, plus /legends-review to run all three with debate.
- Three distinct personas — first-principles/business (Elon), design/DX (Jobs), and engineering rigor (Linus), each reviewing through its own lens and asking its own questions.
- Cross-review debate & consensus —
/legends-review runs solo reviews first, then has the three legends challenge, defend, and concede before producing a unified rating and top actions.
- Actionable reports in three formats — every review writes
reviews/<date>-<reviewer>/report.json (data), report.html (a self-contained interactive dashboard with filters, a done-checklist, and copy-paste fix prompts), and report.md (a diffable checklist for PRs).
- Report merging —
generate_report.py regenerates or merges several solo review JSONs into one combined HTML report.
- One-line install — a
curl | bash installer that drops the skills straight into ~/.claude/skills/, no cloning or config required.
- Eval sets that ship with the skills — every skill carries a versioned eval set (
skills/<skill>/evals/evals.json), and the shared report generator is covered by a golden test (tests/report-format/test.sh). Behavioral claims here are meant to be re-run, not taken on faith.
Install
curl -fsSL https://raw.githubusercontent.com/Atypical-Consulting/claude-legends-review/main/install.sh | bash
That's it. No cloning, no config, no dependencies.
Usage
Open Claude Code in a repo with pending changes and invoke a reviewer:
/elon-review
Or run the full tribunal (requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS):
/legends-review
Each run writes its findings to reviews/<date>-<reviewer>/. Open report.html in a browser for the
interactive dashboard, or regenerate/merge reports from the JSON at any time:
python3 ~/.claude/skills/legends-review/scripts/generate_report.py \
reviews/2026-07-23-elon/report.json reviews/2026-07-23-linus/report.json -o reviews/combined.html
The Reviewers
🔥 Elon Review /elon-review
First-principles thinking. Business value. Speed. Deletion.
"Delete the abstraction layer. Ship without it. If someone screams, add it back."
Reviews through: complexity, performance, security, observability, maintainability, test coverage.
Finds all problems (numbered, with severity). Scales depth to change significance.
Asks: Should this exist? Could it be 10x simpler? What's the opportunity cost?
🍎 Jobs Review /jobs-review
Design obsession. Developer experience. Product vision. Taste.
"The API does what it's supposed to. It just doesn't sing."
Reviews through: API elegance, naming, DX, error messages, documentation, defaults, the last 10%.
Finds all friction points (numbered, with impact). Scales depth to change significance.
Asks: How does this feel? Would you be proud to demo this? Say no.
🐧 Linus Review /linus-review
Engineering rigor. Correctness. Abstractions. Maintainability.
"You've built an abstraction layer that adds complexity without removing any. Delete it."
Reviews through: correctness, abstractions, error handling, security, logging/observability, performance, readability, naming.
Finds all bugs (numbered, with severity). Scales depth to change significance.
Asks: Is this actually correct? What breaks at 3 AM? Show me the code.
🏛️ Legends Review /legends-review
All three. They review independently, then debate in character until consensus.
Step 0 → Framing pass (is this the right question? — before spending three agents)
Phase 1 → Solo review (no groupthink)
Phase 2 → Cross-review & debate (challenges, defenses, concessions)
Phase 3 → Consensus (unified rating, top actions, hard truths)
Produces a final report with:
- Individual ratings and verdicts
- Where they agreed (highest-confidence items)
- Where they fought (and who won)
- Top 5 unified actions
- Three hard truths — one from each legend
Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS enabled in settings.json.
The Actionable Report
Every review — solo or legends — also lands on disk as reviews/<date>-<reviewer>/:
report.json — the findings as data (severity, category, file:line, recommendation, effort)report.html — a self-contained, actionable dashboard generated from the JSON:
done-checklist persisted across reloads, severity/reviewer/category filters,
a copy-paste fix prompt per finding (drop it straight back into Claude Code),
markdown export of what's left, light/dark themereport.md — a diffable checklist for PRs and issues
The HTML is never hand-written — regenerate it anytime from the JSON, or merge several solo
reviews into one combined report:
python3 ~/.claude/skills/legends-review/scripts/generate_report.py \
reviews/2026-07-23-elon/report.json reviews/2026-07-23-linus/report.json -o reviews/combined.html
Sample Output
🏛️ LEGENDS REVIEW — CONSENSUS REPORT
Reviewed by: Elon Musk 🔥 · Steve Jobs 🍎 · Linus Torvalds 🐧
┌──────────┬────────┬──────────────────────────────────────────────────┐
│ Reviewer │ Rating │ Verdict │
├──────────┼────────┼──────────────────────────────────────────────────┤
│ 🔥 Elon │ 7/10 │ "Ship it, but delete the middleware layer." │
│ 🍎 Steve │ 6/10 │ "Someone cared about this. Now care about the │
│ │ │ error messages too." │
│ 🐧 Linus │ 8/10 │ "The error handling is correct. That's rarer │
│ │ │ than you think." │
└──────────┴────────┴──────────────────────────────────────────────────┘
Consensus Rating: 7/10
Where They Agreed:
• Delete the middleware abstraction — it creates complexity without removing any
• Error messages need a complete rewrite for developer experience
• Add integration tests before shipping
Where They Fought:
• Elon wanted to delete the validation layer; Linus insisted it prevents
a class of runtime errors. Linus won.
The Hard Truths:
🔥 "The architecture is fine. The velocity is not."
🍎 "This is a collection of endpoints, not a product."
🐧 "This will break at 3 AM and whoever is on call will curse your name."
Manual Install
git clone https://github.com/Atypical-Consulting/claude-legends-review.git
cp -r claude-legends-review/skills/* ~/.claude/skills/
Why Legends Review?
Most AI code review tools give you one perspective. Legends Review gives you three — and makes them argue.
- Deeper than a linter: Finds security issues, observability gaps, DX problems, and architectural debt that static analysis misses
- Scales to context: Cosmetic changes get a quick pass; new features get the full tribunal
- Exhaustive: Each reviewer finds ALL issues (numbered, with severity), not just the first one
- Checkable: Every skill ships its eval set and the report generator has a golden test — run them yourself
Built for developers who use Claude Code and want more than "LGTM" on their pull requests.
Requirements
- Claude Code CLI
- For
/legends-review: Agent Teams experimental feature enabled
Tech Stack
License
MIT — do whatever you want.
Roadmap
Track progress and proposals in the open issues.
Contributing
Contributions are welcome. Open an issue first to discuss any significant change.
- Fork the repository and create your branch (
git checkout -b feat/my-feature)
- Commit your changes (
git commit -m 'feat: ...')
- Push the branch and open a Pull Request
