菜单
Create subtask document (first half): user confirmation, goals, assumptions. Auto-runs assumption verification experiments and fills in results. Call after Gate 1.1 (user confirmation) is done, before writing any code.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Auto-loop execution workflow with quality gates. Use when starting any non-trivial implementation task. Provides automatic task decomposition, code implementation, testing (L1-L4), and iterative quality gates until completion. Invoke with /autoworker.
Record progress to subtask: Phase completion (from autoworker:code) or test results (from autoworker:test). Auto-detects upstream type from conversation context. Ends by calling autoworker:dispatch.
Implement ONE pending Phase from subtask plan. Only write code, do not mark checkboxes or run tests. Called by autoworker:dispatch. Ends by calling autoworker:checkpoint.
Structured deep discussion for Plan Mode. Runs inside EnterPlanMode to ensure thorough questioning before writing a plan. Covers motivation, assumptions, design, acceptance criteria. Call this immediately after entering Plan Mode for any non-trivial task.
Read subtask.md checkbox state and route to next skill. The ONLY routing point in the execution loop. Called after autoworker:checkpoint, autoworker:gate-check, autoworker:subtask-update, autoworker:subtask-plan, or when resuming after context loss. When lost, call autoworker:dispatch.
Final quality gate before reporting task completion (Gate 3). Fills confidence assessment, runs supplementary verification for <95% items, completes self-check checklist. Writes PASS/FAIL result to subtask.md, then calls autoworker:dispatch for routing.
基于 SOC 职业分类
| name | subtask-init |
| description | Create subtask document (first half): user confirmation, goals, assumptions. Auto-runs assumption verification experiments and fills in results. Call after Gate 1.1 (user confirmation) is done, before writing any code. |
After Plan Mode discussion is complete and user has confirmed direction, invoke this skill to create the first half of subtask.md.
Glob `subtask_*.md` (current directory) + Glob `claude_docs/subtask/*.md` (archive)
→ Extract all sequence numbers → take max + 1
Review the conversation and extract:
Extract assumptions that depend on external systems from plan and conversation. Each assumption needs:
Pure internal logic changes (e.g., refactoring, documentation): Write one line No external dependencies | — | —.
Glob/Grep search for related files → confirm paths and line numbers
Do not guess line numbers — must Read to confirm.
Extract from plan file / conversation:
Glob `subtask_*.md` (exclude subtask_template.md) →
For each file grep `status: active` →
Edit all found to `status: paused`
Purpose: Ensure at most 1 active subtask at any time. Creating a new subtask automatically pauses old ones.
Language rule: subtask.md is an internal work document. Write it in the user's conversation language (as configured in their CLAUDE.md or inferred from their messages). Do NOT use the task's target language — e.g., if the task is "translate to English", the subtask itself should still be in the user's language.
Reference the table structure in subtask template, Write file:
subtask_<sequence>_<name>_<date>.md
First line after title: write status: active (flush left, grep-friendly format).
Include these sections (result columns left empty):
Line by line, read the "verification experiment" column of the assumptions table:
Hard rules:
Subtask first half created: subtask_<sequence>_<name>.md
- Assumption verification: <N>/<N> passed
- Acceptance criteria: <N> (or "No quantitative metrics")
- Paused old subtask: <list paused filenames, or "None">
→ Automatically invoking autoworker:subtask-plan
After outputting the summary above, immediately invoke autoworker:subtask-plan. Do not wait for user instructions, do nothing else.