| name | zero-shot |
| description | Experimental. Aggressively distill LLM-facing prose to imperative essentials. Skills → 1–3 sentences. Procedural skills → numbered list. Multi-section instruction files → bullets per section. |
| model | opus |
| argument-hint | [file-path | skill-name | agent-name | rules] |
| allowed-tools | ["Read","Edit","Glob","Grep","AskUserQuestion"] |
1. Resolve Mode and Target
- Default to dry-run. Use apply mode only if the user asks to change the file.
- Identify target: attached file, pasted markdown, file path, or skill/agent name.
- If target is missing or ambiguous, ask via
AskUserQuestion.
2. Classify
Read file, record word count, then pick classification:
| Classification | Triggers | Output |
|---|
| skill | Single-concept SKILL.md or agent without numbered steps | 1–3 imperative sentences (body only) |
| procedural | SKILL.md or agent with numbered checklist or ## N. headers | Numbered list, 1–3 sentences per step |
| instruction | CLAUDE.md, references/**, rules/**, any non-sequential multi-## doc | Imperative bullets per ## section |
Precedence:
- Common-knowledge override → skill (jokes, greetings, basic formatting, well-known algorithms).
- Numbered checklist or
## N. headers → procedural.
- 2+
## sections, non-sequential → instruction.
- Default → skill.
3. Distill
Aggression: trust LLM priors — never restate well-known formats, algorithms, etiquette, or structure the imperative already implies. Delete-test every sentence: if removal doesn't break the task, drop it. Draft, then cut again — first drafts always under-cut.
Reduction targets (min): common-knowledge ≥95%, skill ≥85%, procedural ≥75%, instruction ≥60%. If draft misses, identify under-performing clauses and cut before reporting.
Write in imperative voice. One verb per bullet, no qualifiers.
Cut:
- Framing (
Overview, Background, Introduction), result descriptions, H1s restating the filename.
- Examples, rationale, illustrative code, redundant "should not" clauses.
- Format restatement after naming the format.
- Audience qualifiers redundant with
clean/safe.
- Error-fallback clauses for common-knowledge tasks — the LLM handles failure.
- Negation pairs where one implies the other.
- Layout instructions implied by a count.
- "Output only X" wrappers when the imperative already scopes the response.
Preserve: frontmatter verbatim, non-obvious code blocks.
Per mode:
- Skill: ≤3 sentences, drop
##. Common-knowledge: one sentence.
- Procedural: numbered list (
1. **Step name.** ...), 1–3 sentences per step, drop ##. Merge adjacent steps sharing one action.
- Instruction: 2–6 imperative bullets per
##, preserve headers and rule/mapping tables.
Output: dry-run prints fenced markdown headed ## Dry Run: <path>. Apply uses Edit.
4. Report
-
Recount before printing. If reduction misses target, return to Distill — never report a miss.
-
Print the table; append (dry run — file not modified) if dry-run.
-
Append ⚠ Experimental — review for lost nuance before committing.
| File | Mode | Before | After | Reduction |
|---|