OOGA BOOGA
Turn verbose SKILL.md -> token-efficient caveman version. Keep all technical substance. Kill fluff.
Goal
Target: ~50-70% token reduction vs original. Floor: skill still usable by agent + readable by human.
Input
User provides:
- File path to SKILL.md
- Pasted skill content inline
- Both (path + content in context)
If path given, read file first. If content in context, use directly.
Caveman Rules (all prose)
Drop:
- Articles: a / an / the
- Filler: just / really / basically / actually / simply / essentially / generally / typically
- Pleasantries: sure / certainly / please / note that / keep in mind / remember to / make sure to
- Hedging: might want to / you may / feel free to / it's worth noting
- Redundant verbs: "in order to" -> "to", "is able to" -> "can", "will need to" -> "need"
- Padding phrases: "at a high level" / "the process of" / "what this means is"
Compress:
- Long noun phrases -> short synonyms: "implement a solution for" -> "fix", "make a determination" -> "decide"
- Common tech terms -> abbreviations: config / auth / DB / fn / impl / req / res / dir / env / ref / doc
- Causality -> arrows: "X causes Y" -> "X -> Y"
- Lists with intros -> just the list (drop "The following are the steps:")
- Passive -> active where shorter: "should be loaded" -> "load"
Preserve exactly:
- All technical terms, method names, file paths, code blocks, command strings
- YAML frontmatter structure (name, description fields — compress description prose)
- Section headers (shorten if unambiguous)
- Numbered/bulleted lists (keep structure, compress text inside)
- Warnings / destructive-action callouts (shorten prose, keep WARNING label)
- All conditional logic ("if X then Y" -> "X -> Y" or "if X: Y")
Fragments OK. Full sentences not required. One word when enough.
Output Structure
Produce complete SKILL.md:
- YAML frontmatter (name unchanged, description compressed)
- Compressed body — same sections, fewer words
- Code blocks / file trees / commands untouched
Do NOT add "this is caveman version" note. Just output file.
Compression Patterns
| Original |
Caveman |
| "Make sure to read the file before editing" |
"Read file before edit." |
| "The following steps should be completed in order" |
(drop — list order implies sequence) |
| "You will need to install the dependencies first" |
"Install deps first." |
| "This is particularly useful for cases where..." |
"Use when..." |
| "It's important to note that..." |
(drop opener, keep the note) |
| "Check whether you have access to X" |
"If X available:" |
| "Feel free to ask the user for clarification" |
"Ask user if unclear." |
| "The user might be asking you to do X" |
"User may want X." |
| "Based on the user interview, fill in these components:" |
"Fill:" |
Process
- Read source skill fully.
- Note: name, core purpose, all required behaviors, file paths.
- Rewrite section by section applying rules above.
- Verify: no tech content dropped. All file paths / commands / code intact.
- Rough token savings check — if <40% reduction, compress harder.
- Output final SKILL.md.
Quality Check Before Output
Ask internally:
- Any file path / command / code block changed? -> revert
- Any required behavior dropped? -> restore terse version
- Any section now ambiguous to agent? -> add clarifying word
- Headers still findable? -> OK to shorten if meaning preserved
Example
Before (excerpt):
Make sure to always read the relevant SKILL.md files before writing any code,
creating any file, or running any other computer tool. This is mandatory
because skills encode environment-specific constraints (available libraries,
rendering quirks, output paths) that aren't in Claude's training data, so
skipping the skill read lowers output quality even on formats Claude already
knows well.
After:
Read SKILL.md before any code/file/tool use. Mandatory — skills encode env constraints (libs, render quirks, output paths) not in training data.
1---2name: ooga-booga3description: Compress SKILL.md into token-efficient caveman version preserving tech accuracy + agent-usability, slashing filler/articles/verbose phrasing. Use when user asks to cavemanify/compress/make caveman/ token-optimize skill, or says "make shorter / fewer tokens / caveman mode". Also trigger when user pastes skill content and asks to compress/minify.4---56# OOGA BOOGA78Turn verbose SKILL.md -> token-efficient caveman version. Keep all technical substance. Kill fluff.910## Goal1112Target: ~50-70% token reduction vs original. Floor: skill still usable by agent + readable by human.1314## Input1516User provides:17- File path to SKILL.md18- Pasted skill content inline19- Both (path + content in context)2021If path given, read file first. If content in context, use directly.2223## Caveman Rules (all prose)2425**Drop:**2627- Articles: a / an / the28- Filler: just / really / basically / actually / simply / essentially / generally / typically29- Pleasantries: sure / certainly / please / note that / keep in mind / remember to / make sure to30- Hedging: might want to / you may / feel free to / it's worth noting31- Redundant verbs: "in order to" -> "to", "is able to" -> "can", "will need to" -> "need"32- Padding phrases: "at a high level" / "the process of" / "what this means is"3334**Compress:**3536- Long noun phrases -> short synonyms: "implement a solution for" -> "fix", "make a determination" -> "decide"37- Common tech terms -> abbreviations: config / auth / DB / fn / impl / req / res / dir / env / ref / doc38- Causality -> arrows: "X causes Y" -> "X -> Y"39- Lists with intros -> just the list (drop "The following are the steps:")40- Passive -> active where shorter: "should be loaded" -> "load"4142**Preserve exactly:**4344- All technical terms, method names, file paths, code blocks, command strings45- YAML frontmatter structure (name, description fields — compress description prose)46- Section headers (shorten if unambiguous)47- Numbered/bulleted lists (keep structure, compress text inside)48- Warnings / destructive-action callouts (shorten prose, keep WARNING label)49- All conditional logic ("if X then Y" -> "X -> Y" or "if X: Y")5051**Fragments OK.** Full sentences not required. One word when enough.5253## Output Structure5455Produce complete SKILL.md:56571. YAML frontmatter (name unchanged, description compressed)582. Compressed body — same sections, fewer words593. Code blocks / file trees / commands untouched6061Do NOT add "this is caveman version" note. Just output file.6263## Compression Patterns6465| Original | Caveman |66| -------------------------------------------------------- | -------------------------------------- |67| "Make sure to read the file before editing" | "Read file before edit." |68| "The following steps should be completed in order" | *(drop — list order implies sequence)* |69| "You will need to install the dependencies first" | "Install deps first." |70| "This is particularly useful for cases where..." | "Use when..." |71| "It's important to note that..." | *(drop opener, keep the note)* |72| "Check whether you have access to X" | "If X available:" |73| "Feel free to ask the user for clarification" | "Ask user if unclear." |74| "The user might be asking you to do X" | "User may want X." |75| "Based on the user interview, fill in these components:" | "Fill:" |7677## Process78791. Read source skill fully.802. Note: name, core purpose, all required behaviors, file paths.813. Rewrite section by section applying rules above.824. Verify: no tech content dropped. All file paths / commands / code intact.835. Rough token savings check — if <40% reduction, compress harder.846. Output final SKILL.md.8586## Quality Check Before Output8788Ask internally:8990- Any file path / command / code block changed? -> revert91- Any required behavior dropped? -> restore terse version92- Any section now ambiguous to agent? -> add clarifying word93- Headers still findable? -> OK to shorten if meaning preserved9495## Example9697**Before (excerpt):**9899> Make sure to always read the relevant SKILL.md files before writing any code,100> creating any file, or running any other computer tool. This is mandatory101> because skills encode environment-specific constraints (available libraries,102> rendering quirks, output paths) that aren't in Claude's training data, so103> skipping the skill read lowers output quality even on formats Claude already104> knows well.105106**After:**107108> Read SKILL.md before any code/file/tool use. Mandatory — skills encode env constraints (libs, render quirks, output paths) not in training data.
cmdecker95