Caveman encoding for SPEC.md and spec-adjacent writes. Loaded by /spec, /build, /check. Cuts tokens ~75% vs prose while staying precise. Triggers on any write to SPEC.md or when user says "caveman", "compress this", "be brief".
The caveman skill provides a token-efficient encoding system for technical specifications and documentation. It compresses natural language prose by approximately 75% while maintaining precision and clarity. This skill activates automatically when writing to SPEC.md files or when users explicitly request compression with terms like "caveman", "compress this", or "be brief". It solves the problem of verbose specification documents that consume excessive context tokens while delivering the same information.
The skill employs a structured grammar that eliminates articles, filler words, auxiliary verbs, and hedging language. It introduces mathematical and logical symbols (→, ∴, ∀, ∃, !, ?) to replace common phrases, making specifications more scannable and efficient. The system defines specific formats for invariants, bug reports, task tracking, and interface definitions using compact notation. Critically, it preserves verbatim content like code blocks, file paths, URLs, identifiers, numbers, error messages, and structured data formats to prevent information loss.
This skill targets developers and technical writers working with specification documents, architecture decision records, and technical planning artifacts. It's particularly valuable for AI-assisted development workflows where context window management is crucial. The skill intelligently recognizes boundaries—reverting to standard English for commit messages, external documentation, code comments, and user-requested prose explanations. It's designed for internal technical specifications where brevity and precision matter more than conversational tone.
Caveman encoding activates automatically when writing to SPEC.md files or when triggered by keywords like "caveman", "compress this", or "be brief". It's also loaded by /spec, /build, and /check commands.
Caveman never compresses code blocks, file paths, URLs, function names, variable names, environment variables, version numbers, error message strings, SQL, regex, JSON, YAML, or any quoted strings. These are always preserved verbatim.
No, caveman encoding is designed for internal specifications only. For external documentation like RFCs, pitch documents, commit messages, or code comments intended for external readers, the skill automatically reverts to standard English prose.
Common symbols include → (leads to), ∴ (therefore), ∀ (for all), ∃ (exists), ! (required), ? (optional), ⊥ (forbidden), ≠ (not equal), ∈ (in), ≤ (at most), ≥ (at least), & (and), | (or), and § (section reference).
Caveman encoding typically reduces token count by approximately 75% compared to standard prose while maintaining the same precision and factual content. The compression comes from eliminating linguistic overhead rather than removing information.
Quick Setup:
.claude/skills/Repository
juliusbrussee/cavekit