スキル一覧に戻る
DazedtilDawn

edge-step

by DazedtilDawn

Mechanical enforcement system for Claude Code - YAML state tracking, hook-based enforcement, Three Gears autopilot

0🍴 0📅 2026年1月21日
GitHubで見るManusで実行

SKILL.md


name: edge-step description: Execute the current step from the plan in active_context.yaml. Use when ready to work on the next task in a structured workflow.

Execute Current Step

Read active_context.yaml to understand the current plan and step.

Instructions

  1. Read the current step from active_context.yaml

    • Find the step matching current_step
    • Understand what needs to be done
  2. Check for reminders - Read .proof/archive.jsonl for recurring weak checks

    • If any check has failed 2+ times across sessions, show a reminder:
    REMINDER: [check_name] has been weak across sessions
    Focus: [specific improvement tip]
    

    Common reminders:

    Weak CheckReminder
    mismatch_detection"Add Expected vs Actual before major operations"
    plan_revision"If this step fails, write a NEW step before retrying"
    tool_switching"If tool fails twice, switch immediately"
    memory_update"After this step, ask: what did I learn?"
    proof_generation"Attach evidence inline, not after"
    stop_condition"If uncertain, frame as bounded options"
  3. Mark it in_progress - Update the step's status in active_context.yaml

  4. Do the work - Execute the step

    • Keep changes minimal and focused
    • If something unexpected happens, STOP and reassess
    • Apply the reminder if one was shown
  5. Verify it worked - Run a test or check

  6. Mark complete - Update status to completed and set proof path

  7. Advance current_step - Increment to next pending step

After Completion

Update active_context.yaml:

current_step: [next step number]
plan:
  - description: "The step you just did"
    status: completed
    proof: "description of evidence"  # or specific artifact path

Add any lessons learned to memory:

memory:
  - trigger: "relevant keywords"
    lesson: "What you learned"
    reinforced: 1

If Blocked

If you cannot complete the step:

  1. Mark status as blocked
  2. Add a note explaining why
  3. Do NOT advance current_step
  4. Report the blocker clearly

Proof Capture (IMPORTANT)

Since Codex CLI doesn't have automatic proof capture, you MUST manually log proof after completing each step.

After Completing Work:

  1. Create proof directory (if needed):

    mkdir -p .proof
    
  2. Append to session log:

    echo '{"timestamp":"<ISO_TIME>","type":"step_complete","step":<N>,"description":"<STEP_DESC>","files":["<FILES>"],"outcome":"success"}' >> .proof/session_log.jsonl
    
  3. Or use the logging skill:

    $edge-log
    

What to Log:

  • Files modified: Actual paths changed
  • Tests run: Commands and results
  • Outcome: success/failure/partial
  • Evidence: Specific proof (test output, line counts, etc.)

Example Log Entry:

{
  "timestamp": "2025-01-15T14:30:00Z",
  "type": "step_complete",
  "step": 3,
  "description": "Add user authentication",
  "files": ["src/auth.py", "tests/test_auth.py"],
  "outcome": "success",
  "proof": "26 tests pass, JWT validation working"
}

This manual logging replaces Claude Code's automatic PostToolUse hook.

スコア

総合スコア

50/100

リポジトリの品質指標に基づく評価

SKILL.md

SKILL.mdファイルが含まれている

+20
LICENSE

ライセンスが設定されている

0/10
説明文

100文字以上の説明がある

+10
人気

GitHub Stars 100以上

0/15
最近の活動

3ヶ月以内に更新がある

0/10
フォーク

10回以上フォークされている

0/5
Issue管理

オープンIssueが50未満

+5
言語

プログラミング言語が設定されている

+5
タグ

1つ以上のタグが設定されている

0/5

レビュー

💬

レビュー機能は近日公開予定です