skill-zipper

Restructure an existing 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, 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 has hard length limits (rules/description-quality.md). measure_tokens.py sizes everything and emits Architecture flags — never ship what the flags call BAD.

Before reasoning about token cost, read rules/progressive-disclosure-model.md — it defines always-loaded vs on-demand and each directory's role; without that model the 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)

Resolve the skill directory path: use it if provided; in direct user mode ask for it; in pipeline mode a missing path is a stage failure to report, not a question to ask.

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 — 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 — Compress/merge/dedup candidates (a references file may duplicate SKILL.md).
• 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: description size, always/on-demand totals, 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 (Harden/Retrigger rewrite, Enrich extraction, plan-listed known-content deletion); 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