메뉴
Apply an LLM judgment task to every in-scope file in the repo, in parallel. Bin-packs files into balanced token slices, fans out general-purpose subagents that read full files, synthesizes results.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Add a deferred entry to the repo's docket — a versioned NN-{slug}.md under docket/{type}/ (tech-debt, feature, follow-up, idea, or upstream), reviewed in the normal PR flow and picked up later. Resolve by deleting the file.
Build a glanceable quest-map of a long session — pure user-intent roots, ground-truth check-marks, [from → to] ranges, PR/docket overlays, a metadata header. A shared recap-on-steroids built by a fresh unbiased agent; light at the top, full detail one cat away. On-demand orientation for when the arc of a long, compacted, multi-thread session is lost and you (and the user) need to decide what to tackle next.
Migrate the binary patcher to a new Claude Code version. Re-derive bytes for changed patches, investigate missing anchors, mark obsoleted patches, verify empirically in fresh sessions, and open a PR. Use when Claude Code releases a new version (announced or stealth-on-CDN) and the patcher needs to follow.
Absorb a new Claude Code version and/or model into this repo — the reusable recipe for what to investigate and how to judge it. Use when CC updates, the model changes, or a loud failure smells like version/model drift.
Draft a PR description into scratch/, pause for your review, then create or update the PR via gh. Gathers git context, indexes the session for semantic search, and follows the repo's CLAUDE.md PR guidelines.
Recover in-flight context from a compacted, crashed, or context-limited session. Indexes transcript for semantic search, reconstructs the session arc.
| name | llm-canvass |
| description | Apply an LLM judgment task to every in-scope file in the repo, in parallel. Bin-packs files into balanced token slices, fans out general-purpose subagents that read full files, synthesizes results. |
| argument-hint | --type|--glob|--all <task> |
| user-invocable | true |
| disable-model-invocation | false |
| effort | max |
| allowed-tools | ["Bash(.claude/skills/llm-canvass/plan-slices.py:*)","Agent"] |
Run a structured LLM judgment task across every in-scope source file in parallel. The user gives the task; the skill handles enumeration, slicing, fan-out, and synthesis.
Run plan-slices.py via the Bash tool — single-quote the task description so
parens, backticks, quotes, and other shell metacharacters in the user's task
pass through unparsed. Flags (--type, --glob, --all, --agent,
--background, --max-agents, --per-agent-tokens) come before the task.
.claude/skills/llm-canvass/plan-slices.py --type py 'your task description'
The companion script:
rg --files under the required scope (--type, --glob, or --all)brief.md and slice-N.txt to the session scratchpadIn a single message, emit N parallel Agent calls. Use the
subagent_type Phase 1 printed (default general-purpose; user override
via --agent). Use the run_in_background value Phase 1 printed
(default false; user opts into background via --background).
For each slice N, the agent prompt is (substitute <scratchpad> with the
absolute path Phase 1 printed on its Scratchpad: line):
Read
<scratchpad>/brief.mdfor your task. Then read<scratchpad>/slice-N.txtfor your file list. Read every file in the list IN FULL. Apply the task.
When all agents return, synthesize the N tool_result blocks into the
shape the task asked for.
--type, --glob, or --all.general-purpose (override --agent <type>).--background).ceil(total_tokens / 400_000), capped at 8.