Caveman Compress

---
name: caveman-compress
description: |
  Rewrite a target file (memory note, doc, scratch plan) in compressed
  caveman form to shrink its token footprint, preserving every technical
  fact, code block, and link. Shows a diff and confirms before overwriting.
  TRIGGER when: /caveman-compress <file> invoked, or asked to compress a doc/memory file.
argument-hint: "<file>"
---

# Caveman Compress

Rewrite a file in compressed form. Read core rules first: `.claude/skills/caveman/SKILL.md`.

## Scope (check first — refuse out of scope)

In-scope targets only:
- files under `memory/**` (notes, daily logs)
- untracked scratch files (not committed to git)
- a file the user passes with explicit "compress this anyway" override

Everything else is **out of scope** — refuse with a one-line reason. In particular:
- `docs/**`, `blog/**`, `README.md`, `CLAUDE.md`/`AGENTS.md` — published / load-bearing prose; brevity is intentional there.
- `tasks/archive/**` and other archives — not read by default; compressing is busywork that mutates the record for ~0 live-token gain.
- `context/**`, `.claude/skills/**`, `.claude/agents/**` — auto-loaded / behavior-steering. Token budget there is `/context-audit`'s job (it has a safe ablation gate); compression here can silently change behavior.

## Procedure

1. **Read** the target file. Refuse politely if no path given, or if the path is out of scope (see above).
2. **Rewrite** prose at `full` level: drop articles/filler, fragments allowed.
3. **Preserve verbatim**: code blocks, commands, URLs/links, tables, frontmatter, file/symbol names, numbers, headings (structure intact).
4. **Show the rewrite** and an estimated token delta. **Confirm before overwriting** — this edits a tracked file; do not write without a yes.
5. On confirm, write the file. Report old → new approx token count.

## Guardrails

- **Never** compress: `LICENSE`, `CHANGELOG.md`, generated/lock files (`*.lock`, `skills.lock`), or anything under `.github/`. These have exact-format contracts.
- **Refuse** files where compression would break a parser (JSON, YAML data, `prd.json`) — compression is for human-prose docs, not machine-read data.
- Compressing a `memory/` note is fine; preserve the `## Heading -- HH:MM UTC` log structure (`context/rules/memory.md`).