aftercare

aftercare — Post-Fix Mistake-to-Skill Capture

Series flow: flirt(explore) → deep-kiss(confirm) → implement → aftercare(restorative care).

After fixing a mistake, or at a session retrospective moment, the user invokes /aftercare. The main context acts only as a thin orchestrator — judgment and drafting are delegated to subagents to keep the main context clean.

When to use

• When the user explicitly invokes /aftercare
• Never trigger automatically

Outputs (3 buckets)

• skill: write a new skill file (SKILL.md + scripts/, references/ if needed)
• memory-file: a short addition to AGENTS.md or CLAUDE.md at the project root
• drop: not worth capturing — do nothing

Orchestration Checklist

Execute in order. Each step is short; the main context handles only dispatch and user interaction.

1. Stage 1 dispatch (candidate extraction + judgment)

1. Prepare the recent conversation context as the dispatch payload. Heuristic: "from the point where the last user intent started, up to the present." When in doubt about how much to include, err on the side of more — include the whole thing rather than trimming.
2. Dispatch a subagent with the body of skills/aftercare/stage1-judge-prompt.md as the prompt and the above context as input.
3. Result schema:

   candidates:
     - topic: <one line>
       bucket: skill | memory-file | drop
       rationale: <one line>
   

2. Present candidates + user selection

• 0 candidates → "No capture candidates found in the recent context." Exit.
• 1 candidate → Show bucket and rationale, then ask "Shall we proceed?" once.
• N candidates → Present a numbered list. "Which would you like to process? (all / number list / cancel)"

3. Process each selected candidate (branch)

drop: One-line notice: "Not recording — reason: ." Move to next candidate.

memory-file:

1. Check whether AGENTS.md and CLAUDE.md exist at the project root → note as project_state.
2. Dispatch a subagent with skills/aftercare/stage2-memory-drafter-prompt.md as the prompt. Input: candidate + project_state.
3. Result schema:

   artifact_type: memory-file
   target_file: AGENTS.md
   operation: append
   section_heading: "## ..."  # omit unless operation=insert-section
   content: |
     ...
   notes: ...
   

4. Show the user target_file, operation, and a content preview, then ask for approval.
5. Approved → if target_file doesn't exist at the project root, create an empty file (one-line notice), then append or insert into the specified section. Rejected with feedback → re-dispatch (2 attempts total). If the limit is exceeded, show the last raw output and ask the user to handle it manually.

skill:

1. Ask the user a scope question:

   "Where would you like to install this skill? (1) project — ./.claude/skills/<name>/ + ./.codex/skills/<name>/ (2) user — ~/.claude/skills/<name>/ + ~/.codex/skills/<name>/"

2. Dispatch a subagent with skills/aftercare/stage2-skill-drafter-prompt.md as the prompt. Input: candidate + rejection feedback if this is a re-dispatch.
3. Result schema:

   artifact_type: skill
   skill_name: <kebab-case>
   files:
     - path: SKILL.md
       content: |
         ...
     - path: scripts/...
       content: |
         ...
   notes: ...
   

4. Show the user the file list, a preview of the first ~30 lines of each file, and notes, then ask for approval.
5. Approved → write files to both the Claude Code path and the Codex path for the selected scope. If .codex/ doesn't exist, mkdir it and give a one-line notice. On filename conflict, ask the user: overwrite / rename / skip.
6. Rejected with feedback → re-dispatch (2 attempts total). If the limit is exceeded, show the last raw output and ask the user to handle it manually.

4. Closing summary

When a single /aftercare invocation is fully processed, print a one-line summary: "Result: skill ×A (path: ...), memory ×B (file: ...), drop ×C."

Errors / Edge Cases

Situation	Handling
Stage 1 output parse failure	Re-dispatch once (with structure-correction instruction). If it still fails, show raw output and ask the user to decide.
Stage 2 output parse failure or user rejection	Up to 2 total retries (attach feedback/error to input). If exceeded: raw output + ask user to handle manually.
Rejection with no feedback	Ask "What should be changed?" once; if no response, skip that candidate.
Filename conflict (skill)	User chooses: overwrite / rename / skip.
.codex/ directory absent	mkdir + one-line notice.

Constraints

• Never trigger automatically — user-explicit invocation only.
• No direct judgment or drafting in the main context — delegate everything to subagents.
• Do not touch ~/.claude/MEMORY.md or the auto-memory system.
• Only AGENTS.md or CLAUDE.md at the project root are valid memory-file targets.