Caveman — token-efficiency skills

Terse commits, PR comments, and optional short chat replies — without compressing kit memory.

Requires pack: caveman. Skills are not on disk until the pack is enabled. See Packs.

npm

npx create-lean-agent-kit@latest . --enable-pack caveman

pnpm

pnpm dlx create-lean-agent-kit@latest . --enable-pack caveman

yarn

yarn dlx create-lean-agent-kit@latest . --enable-pack caveman

bun

bunx create-lean-agent-kit@latest . --enable-pack caveman

What it is

Optional terse output skills adapted from Caveman (MIT) by Julius Brussee. They shrink how the agent writes — commit messages, PR comments, and (optionally) chat replies — without compressing kit memory files.

Caveman is a thin style layer — it does not replace leanagentkit-review, leanagentkit-git-workflow, ambient memory / LEARNINGS.md, or the memory protocol.

Do I need this pack?

• Enable if you want terse Conventional Commits and/or one-line PR review comments (and optionally terse chat).
• Skip if long explanatory replies help you, or you already keep commits short without a skill — terse chat can be net-negative on short Q&A.

Use cases

• Commit formatting — pair leanagentkit-git-workflow (discipline) with leanagentkit-caveman-commit (short message style).
• PR thread comments — leanagentkit-caveman-review for paste-ready one-liners after a full leanagentkit-review.
• Long architecture chats — turn on terse_communication only when replies tend to ramble.
• Team default — commit caveman.yml so everyone shares the same toggles.

What Caveman is / is not

Caveman is	Caveman is not
Terse Conventional Commit message formatting	A replacement for leanagentkit-git-workflow principles
One-line paste-ready PR review comments	A replacement for multi-axis leanagentkit-review
Optional terse agent chat replies	Memory compression for CODEBASE_MAP, ACTIVE_CONTEXT, or LEARNINGS
Opt-in via .leanagentkit/caveman.yml	Always-on guardrails

How it works

Quick start

1. Copy the config example:

   cp .leanagentkit/caveman.yml.example .leanagentkit/caveman.yml

2. Adjust toggles (defaults: commits and reviews on, terse replies off).

3. Re-run stack matching so AGENTS.md §7 lists enabled skills:

   Read .agent/skills/leanagentkit-match-stack.md and follow it.

   Or enable during bootstrap — Step 3e offers Caveman setup.

Config

.leanagentkit/caveman.yml:

enabled: true
terse_communication: false # leanagentkit-caveman — off by default
terse_commits: true # leanagentkit-caveman-commit
terse_reviews: true # leanagentkit-caveman-review

Toggle	Skill	Default	Notes
terse_communication	leanagentkit-caveman	false	Adds ~1–1.5k input tokens/turn when active
terse_commits	leanagentkit-caveman-commit	true	Pairs with leanagentkit-git-workflow
terse_reviews	leanagentkit-caveman-review	true	Pairs with leanagentkit-review

Integration is active only when .leanagentkit/caveman.yml exists with enabled: true. Otherwise skills remain available for explicit invoke — zero impact on the daily loop.

Pairing with LAK skills

leanagentkit-git-workflow     →  commit principles, atomicity, branch discipline
leanagentkit-caveman-commit   →  terse message formatting only

leanagentkit-review           →  five-axis findings report before merge
leanagentkit-caveman-review   →  one-line PR thread comments

leanagentkit-handoff          →  clear prose in HANDOFF.md (not terse)
docs/memory/LEARNINGS.md      →  clear prose for project learnings (not terse)
leanagentkit-caveman          →  terse chat replies only when enabled

Invoke any skill explicitly:

Read .agent/skills/leanagentkit-caveman-commit.md and follow it.

When it helps vs hurts

Helps:

• Long explanatory agent replies (architecture, debugging walkthroughs)
• Formatting commit messages and PR comments where brevity is the goal

Hurts (net-negative):

• Already-terse coding Q&A — skill overhead can exceed output savings
• Compressing kit memory files — use LAK's lean map + distill loop instead

See Caveman honest numbers for measured trade-offs.

Out of scope in Lean Agent Kit

The upstream Caveman repo also ships caveman-compress (memory file rewriting), hooks, stats, subagents, and MCP middleware. LAK does not bundle those — kit memory stays human-readable by design.

Attribution

Adapted from Caveman (MIT) by Julius Brussee.