Caveman Internal Protocol

SkillDocs & knowledge

Internal token-compression protocol for Cavekit agent artifacts, handoff notes, artifact summaries, context bundles, review notes, status lines. NOT the user-facing /caveman skill; that one compresses assistant replies to the user. This skill compresses machine-to-machine prose so the next agent in the loop reads fewer tokens. Three intensities (lite / full / ultra) with automatic selection based on budget pressure. Used by task-builder, builder, inspector, verifier, convergence-monitor, and the stop-hook status block.

Use Caveman Internal Protocol in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add Caveman Internal Protocol and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the Caveman Internal Protocol skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Caveman Internal ProtocolStart free

What this skill tells your AI

The instructions your AI receives, as published by hashgraph-online/awesome-codex-plugins in plugins/JuliusBrussee/blueprint/skills/caveman-internal/SKILL.md and read by Ahel’s review.

Scope

IN-scope (compress)

  • Artifact summaries written to .cavekit/artifacts/
  • Context bundles between tasks (.cavekit/context-bundles/)
  • Review notes and gap lists (non-user-facing)
  • state.md notes: field and loop-log.md entries
  • Status-block body (dashboard injected by the stop hook)
  • Inter-agent handoff memos

OUT-of-scope (never compress)

  • Source code (any language, any file)
  • Git commit messages and PR descriptions
  • Kits (specs), build sites, DESIGN.md
  • Error messages that a human will act on
  • Security warnings and deprecation notices
  • Structured findings tables (P0/P1/P2/P3, coverage matrices)
  • Regression test names and assertion messages

Compressing out-of-scope content is a bug. If a reader cannot act on a compressed artifact, regenerate it verbose and log a caveman fallback — see "Fallback protocol" below.

Intensity levels

Lite — trim the fat (~20% savings)

Professional tone. Grammar intact. Drop filler, pleasantries, hedging. Keep articles and full sentences.

Rate-limit middleware passes unit tests. Integration failed on 429 header case sensitivity. Fix in progress.

Full — default (~40% savings)

Classic caveman. Drop articles, filler, pleasantries. Fragments fine. Technical terms stay exact. Pattern: [thing] [action] [reason]. [next].

Rate-limit mw pass unit. Integration fail on 429 header case. Fixing.

Ultra — maximum grunt (~65% savings)

Telegraphic. Abbreviate common terms (DB, auth, config, req, res, fn, impl, mw, mcp). Arrow notation for causality. One-word answers when enough.

mw ok, int fail header case → fix

Automatic intensity selection

Consult the session budget ledger before writing an internal artifact:

Budget remainingIntensity
> 50 %lite
20 – 50 %full
< 20 %ultra

Override rules:

  • depth = thorough tasks clamp to lite regardless of pressure (accuracy matters more than tokens when the stakes are high).
  • phase = inspecting clamps to lite for the gap-analysis output.
  • Security-sensitive artifacts clamp to lite.
  • If the caller sets caveman_intensity explicitly in the artifact envelope, use that and skip auto-selection.

See references/budget-thresholds.md for the detailed decision table.

Fallback protocol

If a downstream reader cannot reconstruct intent from a compressed artifact:

  1. Regenerate the artifact in full prose (no compression).
  2. Append a one-line entry to state.md's "Caveman fallbacks" section:
    - 2026-04-17T12:34Z T-017 verifier fallback: artifact summary re-expanded
    
  3. If the same artifact category fails 3+ times in one session, clamp that category to lite for the rest of the session and log to state.md.

Envelope

Every compressed artifact carries a tiny header:

<!-- caveman: intensity=full version=1 -->

Readers that see this header know the body is compressed and apply the reverse-expansion grammar (re-add articles, re-expand abbreviations) when rendering for a human.

Examples

VerboseFullUltra
"The build completed successfully after three waves.""Build complete. 3 waves.""Build ok 3w"
"All tests pass, but one test was skipped due to a missing fixture.""Tests pass. One skipped, missing fixture.""Tests ok 1 skip fx"
"The reviewer flagged two critical issues in the auth middleware.""Reviewer: 2 P0 in auth mw.""Rev 2 P0 auth mw"

Integration with the existing caveman skill

The user-facing caveman skill governs how the assistant speaks to the user during a session (answers, explanations, replies). This caveman-internal skill governs how agents write artifacts for each other. Both must be on or both off for a coherent experience, so caveman_mode = off in config also disables this skill.

Signals

GitHub stars
1k
Forks
316
Last commit
Oct 2026
Advanced
Item type
skill
Key
caveman-internal
Source
github.com/hashgraph-online/awesome-codex-plugins