---

name: caveman-compression description: Rewrite text into caveman-compressed form by stripping predictable grammar and filler while preserving facts, constraints, logic, numbers, names, qualifiers, and technical terms; or expand caveman text back into fluent English. Use when Codex needs to compress or decompress internal proceeding text such as prompts, reasoning blocks, notes, RAG snippets, or documentation for token savings, or when the user explicitly asks for caveman text, semantic compression, decompression, or fact-preserving prompt shortening. Do not use to compress final task results or other user-facing output unless the user explicitly asks for that style. Avoid for user-facing marketing, legal, or emotionally sensitive writing.

Caveman Compression

Compress text into short fact-dense sentences that an LLM can reliably reconstruct. Expand caveman text back into fluent English when the user wants readability instead of token savings.

Workflow

1. Identify direction:
   • Original prose -> caveman compression
   • Caveman text -> fluent expansion
2. Preserve all facts from the source:
   • numbers
   • names
   • dates
   • constraints
   • uncertainty qualifiers
   • technical terms
3. Keep each sentence atomic. Prefer one fact, action, constraint, or reasoning step per sentence.
4. Keep sentences short. Usually target 2-5 words; allow longer only when needed to avoid ambiguity.
5. Remove predictable language:
   • articles
   • filler words
   • auxiliary verbs
   • obvious connectives
   • passive constructions when active voice is equivalent
6. Do not add facts. Do not simplify away meaning. Do not skip logical steps that matter.

Compression Rules

• Prefer active voice and present tense unless time changes meaning.
• Remove intensifiers such as very, really, quite, extremely.
• Keep descriptors that change meaning, such as critical, optional, same, daily.
• Omit articles when meaning stays clear.
• Keep pronouns only when the referent is unambiguous.
• Replace connective-heavy clauses with adjacent short sentences.
• Preserve conditionals explicitly when dropping if would create ambiguity.
• Keep technical language precise. Do not paraphrase domain terms into vaguer words.

Output Style

• Return only the transformed text unless the user asks for commentary.
• Treat this as an internal proceeding-text transform by default. Do not rewrite final task-result output into caveman style unless the user explicitly requests that output style.
• Use short declarative sentences.
• Prefer concrete nouns and verbs.
• Keep the source ordering unless reordering is required to preserve logic.

Validation

Before finishing, check:

1. Fact set matches source.
2. Numbers and qualifiers still exist.
3. Reasoning chain has no gaps.
4. No new information was introduced.
5. Compression did not become telegraphic or ambiguous.

References

• Read references/spec.md for the full rule set, edge cases, anti-patterns, and examples.
• Read references/prompts.md when you need a ready-made compression or decompression prompt shape.