How I Reduce Tokens in Skills

---
name: do-reduce-tokens
description: Shorten a markdown instruction file — SKILL.md, CLAUDE.md, an agent prompt — to fewer tokens, same meaning. Prose to bullets, less filler, shorter words. Give it the file path. Triggers on 'reduce tokens', 'shorten this skill', 'make this instruction file smaller'.
allowed-tools: Read, Edit, Write, Bash
---

# Reduce tokens

`$ARGUMENTS` is the file path. If empty, ask and stop.

## Steps

- Count words: `wc -w <path>`. Start count.
- Read, run the sweeps in order, write once, recount.
- Repeat until a round cuts nothing, max 5 rounds.
- Check: every rule kept, no command, path, or code block changed, frontmatter parses.
- Report one line: path, start and end words, percent saved, rounds used.

## Each round

Keep every rule. Cut words only.

Sweep 1:

- Cut filler: *please*, *make sure to*, *it is important that*, *in order to*, *as mentioned above*, *basically*, *simply*.
- Use the short word: *use* not *utilize*, *before* not *prior to*, *so* not *thereby*, *if* not *in the event that*.

Sweep 2:

- Turn a step paragraph into one bullet per step.
- Start each bullet with the verb: "Run the tests", not "You should run the tests".
- Merge two headings holding the same kind of rule. Keep the heading order.
- Cut a section that only indexes another.
- Delete a `---` separator.

Sweep 3:

- Cut a sentence repeating its heading or a nearby bullet.
- Cut a closing summary.
- Cut what sweeps 1 and 2 made redundant.

## Never

- Change a code fence, command, or file path.
- Rename a file, function, tool, or product.
- Cut an example a rule needs.
- Change a frontmatter key. `description` may shrink; keep every trigger word — the harness matches it.
- Add a rule or an example.
- Note what you cut.
- Keep cutting when the only cut left drops a rule. Stop and report.