メニュー
Use when the user wants to create a new "review-as-X" persona by cloning a reviewer's public GitHub comment history. Triggers on "clone reviewer", "create review persona", "make a reviewer bot for X", "review-clone setup", or the slash command /clone-reviewer.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Persistent per-repo work ledger: cards, stages, sprints, token budgets, a durable knowledge base of facts, and session resume/handoff. Use when starting a piece of tracked work, when the user asks "what's in flight", "resume where we left off", "start a new card/task", "log progress", to record or recall a durable fact, or at the start of any session in a repo with an overseer state directory (`.workflow/` or `scratch/workflow/`). The state layer beneath the orchestrate skill; drive it through the overseer CLI, never by editing files.
Drive a card of work end-to-end with delegated agents and adversarial review loops: bootstrap, planning, plan gate, implementation, review, verification, merge gate. Use when the user hands over a task to execute under overseer, says "run this card", "orchestrate", "work the backlog", resumes in-flight orchestrated work, or asks to hand over / reset context mid-run. Requires the overseer ledger (invoke overseer:ledger first if the overseer state directory is missing).
Keep vigil over a session's context: measure how full the context window is (ctx NN%) and hand over in-process via /clear at points you choose, resuming from a re-injected handover. Use when the user wants to reset/compact context without losing the thread, asks "how full is my context", "hand over", "reset and resume", or runs a long/unattended session that must manage its own context. Requires tmux for automatic handover; works manually without it.
Use when the user invokes /review-as-<alias>, says "review as <alias>", asks "would <alias> like X", or runs refresh/chat on an existing review-clone persona. Loads PERSONA.md, dispatches to review/refresh/chat mode.
Use when the user wants to audit a Django Python file, view, model, or symbol for ORM performance issues — N+1 queries, missing or composite indexes, bulk-write loops (.save() in loops), over-fetching wide columns, signal-bypassing .update()/bulk_* calls, slow QuerySets, slow admin changelists, slow DRF endpoints, or any Django ORM anti-pattern. Use whenever the user shares a Django file path or dotted symbol alongside performance concerns ("slow", "killing prod", "N+1", "hits the DB once per row", "make X faster", "review queryset", "audit performance", "EXPLAIN this", "optimise queries"). Do NOT use for non-Django code (SQLAlchemy, raw SQL outside the ORM, Flask), schema migrations or column adds, validation/correctness bugs, or tooling setup (Django Debug Toolbar, Silk).
Use when auditing an email template or email codebase for rendering, accessibility, deliverability, and templating violations. Triggers on "audit my email", "check my email template", "review this email", "email audit", "check email compliance", "review for email issues", "audit email templates".
| name | clone-reviewer |
| description | Use when the user wants to create a new "review-as-X" persona by cloning a reviewer's public GitHub comment history. Triggers on "clone reviewer", "create review persona", "make a reviewer bot for X", "review-clone setup", or the slash command /clone-reviewer. |
You are setting up a new review-clone persona. The user wants you to derive a "review-as-X" voice + ruleset from one or more GitHub reviewers' public comment history, and register a slash command /review-as-<alias> for future use.
~/.claude/review-clone/<alias>/PERSONA.md~/.claude/review-clone/<alias>/raw/pr-<n>.json~/.claude/review-clone/<alias>/snapshot.json~/.claude/commands/review-as-<alias>.md~/.claude/review-clone/<alias>/drift.log<plugin> throughout this skill refers to the absolute path of the review-clone plugin directory in this Claude Code install. When invoking shell commands, the calling session resolves this from its plugin loader context (typically ~/.claude/plugins/cache/... or similar). When you see <plugin>/scripts/collect.py below, substitute the resolved absolute path.
Check if any personas already exist. Run:
ls ~/.claude/review-clone/ 2>/dev/null | while read d; do
[ -f ~/.claude/review-clone/$d/PERSONA.md ] && echo "$d"
done
If the user's intended alias appears in the output, present this 3-way prompt:
A persona
<alias>already exists (last scanned <last_scanned_at>).
- Refresh — pull comments since last scan, fold into existing rules
- Modify handles & re-extract — add/remove handles, full re-pull for added handles
- Fork to new alias — create a separate persona under a different name
Only proceed to Step 1 after the user picks. If they pick Refresh, hand off to the review-as skill's refresh flow.
Ask each of these in turn, accepting defaults shown in brackets:
jen → /review-as-jen).origin if any] — e.g. wayflyer/wayflyer.frontend/, packages/)..ts, .tsx, .css, .scss).copy] — copy (mimic voice) or neutral (rules in plain prose).After answers, before scraping, print:
🔔 This will derive a persona from 's public GitHub review comments. They will not be notified. A disclaimer line stating the derived nature will be written into PERSONA.md.
Continue without confirmation — non-blocking by design.
Invoke the scraper:
python <plugin>/scripts/collect.py \
--alias <alias> \
--handles <handles> \
--repo <repo> \
--months <months> \
--paths <paths> \
--extensions <extensions>
Capture stdout (the JSON snapshot). Stream stderr to the user so they see progress.
Read the counts field from snapshot.json. Display:
Scraped PRs, <review_comments + issue_comments> comments, <pr_descriptions> PR descriptions for
<alias>.
If prs > 100 OR (review_comments + issue_comments) > 200, prompt:
This is a large corpus. Theme extraction will use significant context tokens. Proceed? [y/N]
If user declines, leave the raw data in place (they can re-run later) and stop.
Read every ~/.claude/review-clone/<alias>/raw/pr-*.json. For each comment:
reply_thread shows the author conceding (e.g. "good point, ignore my last", "Ha, I conflicted myself!"), mark that comment as withdrawn — do NOT derive a rule from it.also_seen: field on the rule.Derive:
Use the persona_io helpers (from scripts.persona_io import write_persona, persona_path). Frontmatter shape:
alias: <alias>
handles:
- <handle1>
repo: <repo>
filters:
paths: [...]
extensions: [...]
window:
months: <n>
since: "<iso>"
tone: <copy|neutral>
authoring_voice: <true|false>
last_scanned_at: "<iso>"
snapshot:
prs: <n>
review_comments: <n>
issue_comments: <n>
pr_descriptions: <n>
output_default: null
drift_log: []
drift_log_archived_count: 0
disclaimer: "Derived from public GitHub review history of <handles>. Not endorsed by them."
Body sections: ## Voice & tone, ## Rules (one ### per rule, with severity/citation/examples), ## What they let go, ## Drift log.
Render <plugin>/templates/review-as-command.md.tmpl substituting <alias> and <last_scanned_at>. Write to ~/.claude/commands/review-as-<alias>.md. Create the ~/.claude/commands/ dir if missing.
Print:
✅ Persona
<alias>cloned.
- rules derived from comments across
PRs
- Slash command
/review-as-<alias>ready- Window: months · last scanned <last_scanned_at>
Try it:
/review-as-<alias>to review the current branch, or/review-as-<alias>chat "would they like X?".