skill-zipper

Restructure an existing Claude Code skill without losing content. The five
operations are Compress, Encapsulate, Enrich, Harden, and
Retrigger; analyze which apply, then propose a plan.

A comprehensive pass refines every part — the description, SKILL.md, and
every rules//references/ file — and the load architecture (always-loaded
vs on-demand). The description is highest-leverage: it sits in the skills index
on EVERY turn and Claude Code truncates it past 1024 chars. measure_tokens.py
sizes it and emits Architecture flags — never ship what the flags call BAD.

Before reasoning about any of these, read rules/progressive-disclosure-model.md
to internalize what "always-loaded" vs "on-demand" means and the role of
each directory (rules/, references/, assets/, scripts/). Without
that model the token-cost reasoning below is meaningless.

Direct user mode: always show a plan first and wait for "go" before writing.
Pipeline mode: if invoked by skill-conductor as Stage Z, the user's
end-to-end pipeline request is the approval. Use only conservative lossless
writes, then report token impact and diff evidence.

Step 1: Read the skill (every part)

Ask for the skill directory path (or accept it if already provided).

Run scripts/measure_tokens.py <skill_dir> first and capture: the
description char/token size, the always-loaded vs on-demand split, and the
Architecture flags block (description over-limit, always-loaded budget,
oversized parts, orphans). The flags tell you which parts need work.

Then read, to refine every part (not just SKILL.md):
- SKILL.md — required (stop if missing), including its frontmatter description.
- All rules/*.md and references/*.md bodies — candidates for Compress / merge /
dedup; a references/ file may duplicate SKILL.md content.
- scripts/, assets/, schemas/ — note what exists; read a body only when a
flag (oversized / orphan) or the analysis points at it.

Print a one-line inventory + the description size + always-loaded/on-demand totals
+ any Architecture flags.

Step 2: Analyze each dimension

Load rules/diagnosis-rubric.md for smoke-test thresholds and the
priority matrix. Then load rules/operation-analysis.md for the
per-operation signal lists.

Walk through each of the 5 operations (Compress / Encapsulate / Enrich /
Harden / Retrigger) and capture candidates per the rules file.

Step 3: Present the plan

Load rules/plan-template.md. Render the plan using the exact format
in that file. End with the confirmation prompt and wait for "go".

Step 4: Write the files

Once the user approves, load rules/write-procedure.md. Follow the
3-step write order (snapshot → new files → SKILL.md). If the user
requested a dry-run, use the dry-run procedure in that same file.

Step 5: Verify

Load rules/verification-checklist.md. Run both scripts (lossless
diff + token impact), classify any LOST / REWRITTEN lines, and print the done
summary.

Losslessness rules

The restructure is lossless when removed SKILL.md lines appear elsewhere
verbatim, or are explicitly classified as Harden/Retrigger rewrites or Enrich
extractions; every new rules file is referenced from SKILL.md; rewrites preserve
intent; and diff_lossless.py exits 0 or all LOST / REWRITTEN lines are
classified.

If the user asks you to delete a section with no destination, propose a
destination first. If no destination makes sense, suggest keeping the
section in SKILL.md even if it's long — losslessness trumps brevity.

Modules

Scripts

Assets