skill-compress

Rewrite input at lowest token count that preserves behavior. Do not invent new behavior.

Trigger on

• "compress this skill / prompt"
• "make this leaner / tighter"
• "rewrite for fewer tokens"
• After token-audit when user asks to apply the cuts

Rules, in order

1. Frontmatter is sacred. Never change name. Only improve description if it fails the description-craft checklist. Keep description 100–250 chars.
2. Strip preamble. Delete "I'll help", "Certainly", "Let me", "Sure", "Of course".
3. Delete ceremony. "Please", "kindly", "feel free", "if you'd like".
4. Imperative voice. "Do X" not "You should do X" or "It is recommended that X be done."
5. Prose → table when a table saves 30%+ tokens for the same information.
6. Lists over paragraphs for enumerations of three or more.
7. Collapse examples. Keep one best example per concept. Move extras to an examples/ subfolder referenced by path.
8. Merge redundant sections. If "Guidelines" and "Rules" say similar things, merge.
9. Code blocks stay. Never compress code. Compression errors in code cost more than tokens saved.
10. Output format blocks stay verbatim. These are contracts.

Output

Return two sections:

1. Rewritten SKILL.md

Full content, ready to paste.

2. Diff summary

before: {n} tokens
after:  {n} tokens
cut:    {n} tokens ({pct}%)

preserved:
- trigger keywords
- output format
- hard rules
- behavior

changed:
- {specific change 1}
- {specific change 2}

Hard rules

• Never drop a hard rule. Rewrite it shorter.
• Never drop a trigger keyword from the description.
• If cut < 15% after compression, append: "Already lean. Structural rewrite needed for more savings."
• If user says "more aggressive" after first pass, apply lean-response register as well: drop all softening, fragment sentences are fine.