Cut

Rewrite a skill's prose into the fewest words that steer the model as far, for a skill built elsewhere or one the user wants tightened. The bar is canon.md and lenses/lens-leanness.md; this mode applies it file by file and checks that no rule was lost. A skill Smithy writes gets the same pass in ship.md. Its outcome: every prompt file of {target} rewritten or left as already lean, every rule, convention, reference, protected passage and decision intact, nothing committed, and the user reading the diff beside the before and after table.

Protected passages

Set {target} to the skill folder. Protected by default: each file's frontmatter, which holds the description; an agent's customize.toml, which holds its persona; and what a shape template put there, a rendered skill's SKILL.md and the activation section of its workflow.md. cut_check.py reports a changed frontmatter itself; pass the rest to it as --protect <file> or --protect <file>#<heading>. Ask once what else to protect: a passage kept for its voice, a whole file or section, a decision. When the work on this skill comes from a ticket, the Decision: lines in its Notes are decisions every cutter keeps.

Plan

Copy {target} to {reports_folder}/<YYYYMMDD-HHMMSS>-<skill>-before/; that copy is the old version every later step reads. Count it with uv run {skill-root}/scripts/count_tokens.py <copy>. The prompt files are the .md files outside scripts/ and the root customize.toml; leave protected files and those under about 150 tokens uncut. Show the plan in chat: each file with its tokens, the files not cut, the protected passages, the decisions. Then wait for the user's go.

Examples

Before the cutters run, take each worked example (a sample artifact, a filled-in template, a transcript) through test 8 in canon.md with the user, settling a doubtful one with a fresh subagent given the rule without the example, and make the agreed edit yourself. Cutters left to it keep examples in.

Rewrite and check

Launch one subagent per file to cut, in parallel, with the Cutter prompt from modes/cut-roles.md, filled with the file, the protected passages in it, and the decisions. Then run uv run {skill-root}/scripts/cut_check.py <copy> {target} [--protect ...] and restore, from the copy, every item it lists under a file's lost and every file under missing. Then one subagent per rewritten file runs the Checker. Restore each loss and meaning change it lists as the shortest sentence that states it, and apply the shorter wordings you agree with.

Run the list under Checks in ship.md and fix until clean. When SKILL.md calls render_skill.py, render the skill with uv run --no-cache {project-root}/_bmad/scripts/render_skill.py --project-root {project-root} --skill {target}, once as shipped and once per allowed value of each selector with --set workflow.<key>=<value>; each render finishes without error. Where that script does not exist, report the render as not run.

Hand over

Run cut_check.py once more and restore anything it lists. Run count_tokens.py on {target} and give a table in chat: file, tokens before (from the copy), after, percent, and the total. Under it list the Cutters' doubts and any loss you chose not to restore, with why. Commit nothing; the user reads diff -ru <copy> {target} and commits what they accept.

Without subagents, run each role inline, one file at a time, and say so; read the old version again before checking your own rewrite, and name your judgment on test 8 as judgment.