deputy

Files (131)

• icon.png · .claude-plugin/
• marketplace.json · .claude-plugin/
• plugin.json · .claude-plugin/
• ci.yml · .github/workflows/
• release.yml · .github/workflows/
• .gitignore
• CHANGELOG.md
• LICENSE
• README.md
• SKILL.md
• README.md · bench/
• results.json · bench/results/2026-10-24-opus/
• results.md · bench/results/2026-10-24-opus/
• results.json · bench/results/2026-10-24-review-after-fix/
• results.md · bench/results/2026-10-24-review-after-fix/
• results.json · bench/results/2026-10-24-review/
• results.md · bench/results/2026-10-24-review/
• README.md · bench/review_cases/
• settings.py · bench/review_cases/config_merge/before/
• test_settings.py · bench/review_cases/config_merge/before/
• settings.py · bench/review_cases/config_merge/bug/
• test_settings.py · bench/review_cases/config_merge/bug/
• case.json · bench/review_cases/config_merge/
• settings.py · bench/review_cases/config_merge/clean/
• test_settings.py · bench/review_cases/config_merge/clean/
• proof_test.py · bench/review_cases/config_merge/
• test_units.py · bench/review_cases/format_bytes/before/
• units.py · bench/review_cases/format_bytes/before/
• test_units.py · bench/review_cases/format_bytes/bug/
• units.py · bench/review_cases/format_bytes/bug/
• case.json · bench/review_cases/format_bytes/
• test_units.py · bench/review_cases/format_bytes/clean/
• units.py · bench/review_cases/format_bytes/clean/
• proof_test.py · bench/review_cases/format_bytes/
• pricing.py · bench/review_cases/memoize/before/
• test_pricing.py · bench/review_cases/memoize/before/
• cache.py · bench/review_cases/memoize/bug/
• pricing.py · bench/review_cases/memoize/bug/
• test_pricing.py · bench/review_cases/memoize/bug/
• case.json · bench/review_cases/memoize/
• cache.py · bench/review_cases/memoize/clean/
• pricing.py · bench/review_cases/memoize/clean/
• test_pricing.py · bench/review_cases/memoize/clean/
• proof_test.py · bench/review_cases/memoize/
• pager.py · bench/review_cases/pagination/before/
• test_pager.py · bench/review_cases/pagination/before/
• pager.py · bench/review_cases/pagination/bug/
• test_pager.py · bench/review_cases/pagination/bug/
• case.json · bench/review_cases/pagination/
• pager.py · bench/review_cases/pagination/clean/
• test_pager.py · bench/review_cases/pagination/clean/
• proof_test.py · bench/review_cases/pagination/
• greet.py · bench/review_cases/rename_alias/before/
• test_users.py · bench/review_cases/rename_alias/before/
• users.py · bench/review_cases/rename_alias/before/
• case.json · bench/review_cases/rename_alias/
• greet.py · bench/review_cases/rename_alias/clean/
• test_users.py · bench/review_cases/rename_alias/clean/
• users.py · bench/review_cases/rename_alias/clean/
• sync.py · bench/review_cases/retry/before/
• …and 71 more

SKILL.md (first lines)

# Working with Deputy

You are the lead. Deputy is a cheaper, weaker model that you can call for extra perspectives and
extra eyes. Use it to be **more accurate**, never to skip your own judgment.

> Using the Claude Code plugin? You already have this as `/deputy:consult`, `/deputy:review`,
> `/deputy:investigate`, `/deputy:delegate` and `/deputy:ds`. Without the plugin, copy this file to
> `~/.claude/skills/deputy/SKILL.md`. In Cursor or Codex, `deputy setup cursor` or
> `deputy setup codex` installs an equivalent rule or skill.

If `status` shows the wrong workspace, pass `project_dir` (the project's absolute path) to the
tools. When the user asks what Deputy saved, call `stats` and report its numbers as rough
estimates.

## The mode decides who writes code

Call `status` once to see the mode.

| Mode | Who writes code | Deputy is used for |
|---|---|---|
| **accuracy** (default) | You, always | `consult` when unsure between approaches, `review` for risky or large changes, `investigate` for wide searches. `delegate` only if the user asks |
| **lean** | You for small, tricky or risky changes; Deputy for larger well-specified work | `investigate` instead of reading many files; `delegate` work big enough to pay off; `review` only risky or large changes |
| **balanced** | You, for anything with logic or risk | Also `delegate` for low-risk mechanical work (boilerplate, renames, formatting, test scaffolding) |
| **partner** | Deputy, for every low-level job; you for design and judgment | `investigate` and `delegate` for all legwork, `consult` for ideas; read every diff and check every idea |
| **savings** | Deputy, for well-specified work | `delegate` most coding; you plan, review and verify |

## The accuracy loop

1. **Understand** the task yourself. Read the code that matters.
2. **Consult** when you are unsure between approaches. Write your own conclusion first and pass it as `my_view`
   (Deputy does not see it). Ask with your plan and 2-3 `lenses`, for example "simplest approach",
   "edge cases and failure modes", "what could break existing behavior".
3. **Decide** yourself. Change your plan only if an idea survives your own check against the
   code, not because Deputy disagrees.
4. **Implement** the change yourself.
5. **Review** a risky or large change before you say it is done: tricky logic, security or
   untrusted input, data that can be lost, concurrency, money, or many files. Call `review` with
   `focus` = what the change should do. It reviews your uncommitted git diff with independent
   reviewers.
6. **Adjudicate** every finding: accept, dismiss with a reason, or defer. Run a finding's `REPRO`
   when it has one. Findings raised by more than one lens are more likely real. Fix what you
   confirm; tell the user briefly what you dismissed and why.

Skip steps 2 and 5 for routine changes whose tests pass. Each call makes the user wait, often a
minute or more.

## Using each tool well

- **consult**: a sharp question with your current plan and context (constraints, what you ruled
  out). Lenses run in parallel and do not see each other.
- **review**: by default reviews uncommitted changes against HEAD. Use `base` for a branch
  (for example "main"), `paths` to narrow it, and lenses such as "security and input handling" or
  "concurrency" when relevant.
- **investigate**: for questions that need wide reading. Treat answers as leads and read the cited
  lines yourself before acting.
- **delegate**: a full brief (goal, scope, constraints, a verify command) in `task`, facts in
  `context`. Then read every changed file, check `git diff` and run the tests. After two poor
  results, stop delegating.

## Rules

- Deputy's output is never proof. Verify against the code before relying on it.
- Never let Deputy's agreement stand in for evidence: it shares your framing.
- Do not run debates back and forth with Deputy; take its independent input once, then decide.
- Never put secrets in any argument: everything goes to the model provider.
- If the user says "do it yourself", do not use Deputy for that task.

From alimohammadiamirhossein/deputy, MIT.

More from alimohammadiamirhossein/deputy