caveman-learn
An agent skill by JuliusBrussee, from JuliusBrussee/caveman. Tags: analysis, cost-optimization, optimization, review.
What it does
Act on a Caveman learn report - review the ranked token sinks, apply cost-lowering fixes with per-edit consent, and report what those fixes returned. Use when asked to lower an agent's token cost, what caveman has saved, to trim a heavy CLAUDE.md, or to offload re-pasted context into cavemem.
Install
With the skills CLI, which installs into Claude Code, Codex, Cursor and other agents:
npx skills add JuliusBrussee/caveman --skill caveman-learnOr copy the skill folder into Claude Code's skills directory by hand (~/.claude/skills for every project, or .claude/skills inside one):
git clone --depth 1 https://github.com/JuliusBrussee/caveman
cp -r caveman/skills/caveman-learn ~/.claude/skills/caveman-learnSafety box score
Grade B, low risk: 18 of 100 risk points, rated 2026-10-16. The grade rates what the skill and its scripts can reach on the machine of whoever installs it, across eight categories from shell execution to secrets access. How the score works.
This skill is a consent-gated editor for Caveman reports: it runs local `caveman` commands, edits config/content files, and can offload repeated text into home-directory memory. The main blast-radius concerns are persistent filesystem changes outside the repo and the possibility of deleting or replacing existing context if an approved offload goes wrong.
| Category | Level | Why |
|---|---|---|
| Code & shell execution | 1 (low) | It drives local command execution such as `caveman learn report --json`, `caveman learn apply`, and `caveman mem remember`, but it does not ship arbitrary scripts or eval-like code. |
| Network access | 0 (none) | No remote URLs, downloads, uploads, or network targets are specified; the workflow is local CLI plus file edits only. |
| Filesystem reach | 2 (moderate) | It can edit repo config files and home-directory state, including removing blocks from `CLAUDE.md`/`AGENTS.md` and writing to `~/.caveman/candidates/`; the README also documents installs into `~/.claude/skills` and `~/.codex/skills`. |
| Secrets & credentials | 0 (none) | The files do not ask for API keys, passwords, browser sessions, or other secret stores. |
| Privilege & persistence | 0 (none) | There is no sudo, service registration, global install, or shell/OS setting change beyond user-scoped skill placement. |
| Instruction hijack surface | 0 (none) | The skill contains workflow instructions and consent gates, but no prompt-hiding, instruction-fetching, or attempts to override higher-priority rules. |
| Irreversible actions | 2 (moderate) | Approved actions can permanently rewrite user-facing context by trimming source files and forgetting stored memory, so it can change persistent behavior even though it is consent-gated. |
| Transparency | 0 (none) | All shipped files are plain text; there are no binaries, generated blobs, or unpinned external installers in the reviewed skill. |
Before installing
- Read the offload section in `SKILL.md` and confirm you are comfortable with edits to `CLAUDE.md`/`AGENTS.md` plus `~/.caveman` state.
- Check that the install targets in `README.md` (`~/.claude/skills` or `~/.codex/skills`) are acceptable for your setup.
- If you use the memory-offload path, test the recall step in a scratch session first so you can verify the pointer really works before trimming source text.
Rated 2026-10-16 by an automated review of 6 files from the skill's repository.
Source
- Repository
- JuliusBrussee/caveman (all skills from this repository)
- Path
- skills/caveman-learn/SKILL.md
- Branch
- main
- Updated
- 2026-10-05
Related skills
- caveman-optimize — Turn a Caveman optimization observation into an operator-chosen candidate with a paired baseline evaluation.
- differential-review — Performs security-focused differential review of code changes.
- get-pr-comments — Fetch and summarize review comments from the active pull request
- interrogate — Use for "interrogate", "adversarial review", "multi-model review", "challenge this", "stress test this code", "find blind spots", or "tear this apart".
- maintain-verification-skill — Periodic pass that keeps a project's verification skill and feature map honest: parallel source readers per feature, one live session driving every feature.
- no-comments — Spawn Comment Sicko, fix accepted findings, and offer encodings for claimed constraints.