| name | caveman-ify |
| description | Compress any skill to minimal tokens — rewrites SKILL.md, all references/, and scripts/ comments. Strips verbose prose, filler, redundancy while preserving technical accuracy and functionality. Use when user says "caveman-ify this skill", "compress skill", "shrink skill", or invokes /caveman-ify.' |
Caveman-ify
Take existing skill. Make small. Keep smart.
Inputs
- Path to skill folder (or skill name if in known location)
- Compression level: standard (default) or aggressive
If no path given, ask.
Workflow
- Read entire skill:
SKILL.md, all references/*.md, all scripts/ file comments
- Show token estimate before/after (rough line count as proxy)
- Rewrite everything. Wait for user confirmation before writing files.
Compression Rules
Apply to ALL files in skill (SKILL.md, references, script comments):
Drop:
- Articles (a/an/the)
- Filler (just/really/basically/actually/simply/note that/it's worth noting/keep in mind)
- Pleasantries (sure/certainly/of course/happy to)
- Hedging (might/perhaps/possibly/consider/you may want to)
- Redundant examples when one suffices
- Restating what was already said
- "For example" / "In other words" / "That is to say"
Keep exact:
- Technical terms, API names, flags, paths
- Code blocks (unchanged)
- Error messages (quoted exact)
- Frontmatter
name and description fields (rewrite description to be terse but keep discovery triggers)
- Constraint values (char limits, regex patterns, etc.)
- Decision logic / branching conditions
Transform:
- Long sentences -> fragments
- "You should X because Y" -> "X. Reason: Y."
- "In order to" -> "To"
- "Make sure to" -> drop, just state the action
- Bullet prose -> terse bullets
- Multi-sentence explanations -> single line where possible
- Paragraphs -> bullets or tables
Aggressive mode (additional):
- Abbreviate common terms (DB/auth/config/req/res/fn/impl/repo/dir/msg)
- Use arrows for causality (X -> Y)
- Merge related sections
- Strip all examples except one per concept
Output
Rewrite files in-place (after confirmation). Show before/after line counts.
Guardrails
- Never alter script logic, only comments
- Never change frontmatter
name
- Never remove decision gates or branching conditions
- If skill has scripts, validate they still parse after comment compression:
node -c <file>
- Preserve all file paths and cross-references between files
