メニュー
Research codebase and create spec + phased implementation plan with file-level tasks
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
| name | plan |
| description | Research codebase and create spec + phased implementation plan with file-level tasks |
| triggers | ["plan","spec","battle_plan"] |
| priority | 10 |
| keywords | ["plan","spec","feature","bug","refactor","implement","design","breakdown"] |
WORKFLOW:
# Specification: {Title}
**Track ID:** {trackId}
**Type:** Feature|Bug|Refactor|Chore
**Created:** {date}
## Summary — 1-2 paragraphs from research
## Acceptance Criteria — 3-8 concrete checkboxes
## Dependencies
## Out of Scope
## Technical Notes — architecture decisions, reusable code found
# Implementation Plan: {Title}
**Spec:** [spec.md](./spec.md)
## Phase 1: {Name}
### Tasks
- [ ] Task 1.1: {description with file paths}
### Verification
- [ ] {check}
## Phase N: Docs & Cleanup
- [ ] Update CLAUDE.md
- [ ] Remove dead code
## Context Handoff
### Key Files — from research
### Decisions Made — why X over Y
### Risks
RULES:
WRONG: Plan without reading code first (guessing file paths) CORRECT: research → spec → plan → update_plan → finish
EXAMPLE — Plan a new feature: Instruction: "plan adding WebSocket support"
Self-improving optimization via Karpathy autoresearch pattern. Generates → evaluates → scores → mutates prompts/descriptions in a loop. Targets — tool-selection, system-prompt, skill, decision-parser. Use when "optimize tools", "autoresearch", "improve skill X", "self-improve prompts", "optimize tool descriptions".
End-to-end project improvement — plan, branch, delegate execution, verify. Calls /plan for analysis, then delegates fixes to agents. Use when "improve project", "swarm improve", "send agents to fix", "refactor project X", "upgrade project". Do NOT use for planning only (use /plan), single tasks (use /delegate), or own codebase (use /self-evolve).
Autonomous task loop — iterate on a task until done, commit after each step, run tests. Use when "run in loop", "autonomous mode", "keep working", "don't stop", "bighead mode", "--loop". Do NOT use for one-shot tasks (just run normally) or self-improvement (use /self-evolve).
Discover and install agent skills from skills.sh (60K+ skills). Use when "find a skill", "search skills", "is there a skill for", "install skill", "skill catalog". Do NOT use for auditing existing skills (use /skill-audit) or browsing local skills (use `skills list`).
Self-improvement mode — analyze past sessions for recurring failures, patch own code, test, commit, rebuild. Use when "evolve", "self-improve", "fix yourself", "analyze sessions", "--evolve". Do NOT use for normal coding (use /bighead) or delegating tasks (use /delegate).
Orchestrate multiple CLI agents in parallel — delegate tasks to claude/gemini/codex/opencode/rust-code via .tasks/ system. Use when "delegate to", "spawn agents", "parallel analysis", "send task to claude", "run in background", "orchestrate agents". Do NOT use for in-process sub-agents (use spawn_agent) or single bash commands (use bash_bg).