loop-distiller

Loop Distiller

Turn real session history into a pre-filled loop-builder blueprint. This skill is the evidence front-end; loop-builder is the interview and the single scaffolding authority. Output is always a loop — never a standalone skill.

The backbone knowledge is references/distillation.md (rubric + blueprint + handoff). Load it when you run the flow.

When this fires

On demand only (Trigger A). The user asks to distill what they just did. There is no background watching in this phase — a cross-session scanning loop (Trigger C) is designed in the spec but not built yet.

Procedure

1. Read the transcript. Prefer the current session already in your context. For precision or after a /clear/compaction, re-read it deterministically:

   python3 -c "import sys; sys.path.insert(0, '${CLAUDE_PLUGIN_ROOT}/skills/loop-distiller/scripts/distiller'); \
   import transcript_reader as tr, json; p = tr.locate(); \
   print(json.dumps(tr.condense(tr.parse(p)), indent=2) if p else 'NO_TRANSCRIPT')"
   

   The script's transcript-path scheme is an observed Claude Code convention — verify against current docs; if it returns NO_TRANSCRIPT, fall back to the in-context history.

2. Apply the rubric from references/distillation.md. Count how many of the seven blocks have evidence. Lenient gate: 4+ blocks → proceed. Fewer → tell the user this session doesn't look like a reusable workflow yet, and stop. (Saying no is the point — most sessions are one-offs.)

3. Draft the blueprint. Emit loop-builder's fill-in template with each line drawn from evidence, an ⟵ ev: cite, and a confidence marker. Any block without evidence is UNKNOWN → ASK. Never fabricate a value or a /goal /loop /schedule flag — flag uncertainty instead.

4. Hand off to loop-builder. Present the draft for correction, then invoke the loop-builder skill with the corrected answers as its interview starting point — the corrected answers are passed by inlining them in the conversation, not via a file or argument contract. Do not scaffold loop files here.

What this skill must not do

• Never write a loop, a SKILL.md, a STATE.md, or any scaffold file — that is loop-builder's job.
• Never emit a standalone skill (no Hermes-style skill sprawl — loops only).
• Never invent evidence. UNKNOWN → ASK beats a confident guess.

Collecting feedback and reporting bugs

On an error or when clearly blocked, capture locally (never auto-filed):

python3 "${CLAUDE_PLUGIN_ROOT}/skills/feedback-to-issue/scripts/feedback/cli.py" \
  append --category bug --text "<what broke + context>"

To review and file, invoke the loop-builder:feedback-to-issue skill.