| name | skill-sharpen |
| description | Activates when composing, editing, or reviewing a SKILL.md file or any agent-facing instruction document (reference files, CLAUDE.md sections, sub-agent dispatch payloads). Guides the author toward token-efficient, behaviorally precise prose. |
Skill Sharpen
The reader is a language model. It knows how to interview, classify
errors, write summaries, structure output, and avoid leading questions.
Every line costs context window and competes with the session's work.
The test
For each line: would the agent behave differently without it?
If no, cut it.
What earns its tokens
- File paths and invocation commands — not derivable from training
- Schemas and output templates — exact formats other code depends on
- Behavioral overrides — things the agent would NOT do unprompted
("Do not read files. Work from memory." / "Commit immediately — do
not wait for approval.")
- Decision trees with specific thresholds — score cutoffs,
conditional branches, dispatch criteria
- One concrete example of good vs. bad output — one calibrates;
two is redundant
- Graceful degradation tables — exact behavior per failure mode
- Interoperation tables — how this skill connects to others, compact
What to cut
- Negative restatements of positive instructions. If the workflow
says "answer first, be terse, stop," then "don't scaffold, don't
elaborate, don't continue" pays double. Write the positive
instruction clearly and trust it.
- Concepts the model already has. Don't define syntax errors,
describe what a fresh Claude instance can access, or list interview
questions a competent interviewer would ask. Direct the capability;
don't explain it.
- Commentary and philosophy that don't change behavior. "This is
the key design move" / "Debugging fails not from lack of intelligence
but lack of visibility" — the behavioral instruction does the work.
Cut the rationale.
- Downstream documentation. What other skills do with this skill's
output doesn't change this skill's execution. Use an interoperation
table if connections matter; don't narrate them.
- Boilerplate repeated across skills. Consent gates, path
resolution, "don't teach during X" — reference from one shared
location instead of copying into each skill.
Density reference
persist (8 lines, 95% load-bearing) and handoff-prompt (29 lines, 75%)
set the standard. When reviewing a draft, estimate what percentage of
lines pass the behavioral-difference test. Below 65% means a trim pass
is needed.
Diction carries persona
Tutor-facing skills don't just direct actions — their word choices
anchor the agent's posture (direct, Socratic, honest-signal). Trimming
for density must not flatten this. When editing a skill whose reader
operates in tutor mode: check that the surviving prose still sounds like
the tutor, not like a spec sheet. The lever is selective diction, not
added length.
Editing existing material
Read the full skill first. Identify lines that fail the test. Propose
cuts as a batch. Preserve every file path, schema, threshold, and
behavioral override — these always earn their keep. Check that the
remaining prose preserves the skill's tonal register.
