Compress Skill

Shrinks an existing SKILL.md into a short, decision-critical core, relocating dense/large/redundant content into reference files. Preserves every rule, threshold, and edge case — nothing is lost, only relocated or deleted if truly redundant.

Steps

#	Step	Action	Output / Gate
1	Read target	Read the full SKILL.md	inventory of sections: steps, branching logic, tables, verbatim templates, Notes/recap
2	Read existing refs	Read every file already linked from the skill (references/*.md or similar)	flag any that describe a stale/older mechanism than the current main file — these must be rewritten, not just kept
3	Classify content	See classification table below	each section tagged: table / flowchart / extract-to-ref / delete
4	Design ref file set	One file per template category; one file for detailed mechanism/flowchart; keep existing example/troubleshooting files (rewrite if stale per step 2)	file list finalized before writing
5	Write reference files	Write tool, refs first	main file's pointers can be one line each since detail lives here
6	Rewrite main SKILL.md	See rewrite checklist below	short core file + "Reference files" map at the end
7	Verify nothing lost	Diff old content section-by-section against new tables/refs	every rule/threshold/edge case appears somewhere in the new set
8	Report	State size delta + what moved where	e.g. "298 → 55 lines; templates → agent-prompts.md, flowchart → orchestrator-model.md, stale example rewritten"

Classification table (Step 3)

Content shape	Treatment
Linear numbered steps with input/output	table: step, input, output/gate
Branching logic (if/else, retry/skip/abort)	table (condition → action); use mermaid flowchart instead if branches loop/cycle
Verbatim templates (agent prompts, code payloads)	extract to dedicated reference file, one-line pointer left in main file
Notes/recap section restating the body	delete — redundant content has zero compression floor, don't summarize it, remove it
Wordy prose inside an already-tabular cell	trim in place — see wording-tightening pass below; no restructure needed

Wording-tightening pass (already-tabular content)

When a table/section already exists but cells are verbose, tighten wording without dropping substance:

• Drop filler: "in order to" → "to", "make sure to" → drop, "you should" → drop, restated rationale already implied by a command.
• Cut connective throat-clearing ("Also actually open...", "Note that...") — keep the instruction, drop the preamble.
• Shorten to fragments: commands, thresholds, flags, and named tools stay verbatim; surrounding prose becomes clipped phrases.
• Collapse repeated qualifiers ("don't X, and don't Y, and don't Z" → "don't X/Y/Z" or a short list).
• Keep every number, flag, command, filename, and named exception — only the connective tissue is cut.
• Apply per-cell, not per-file: a cell can be tightened even when the surrounding table structure is otherwise left alone.

Rewrite checklist (Step 6)

• Keep frontmatter (name, description, metadata) unchanged unless scope changed — bump version if content meaningfully changed.
• Open with 2-3 sentences: who drives what, the one non-obvious rule (a gate, a threshold, an ownership boundary).
• Replace numbered prose steps with the Step 1 table's output.
• Replace branching prose with condition→action tables or a flowchart.
• Replace verbatim templates with one-line pointers to extracted files.
• Delete sections that only recap content already covered elsewhere.
• End with "Reference files" list — one line per file, stating what it contains.

Reusable heuristic

Compress only what is dense and unique. Delete what is redundant. Relocate what is large and copy-paste (templates). Keep in the main file only what is decision-critical and short (tables, gates, thresholds).