Skill Pruning

The final pass of the Skill Authoring Checklist: a smallness discipline framed as hunting a set of failure modes that bloat a skill. Massive skills are treated as a symptom of one of these, not a root cause, so pruning is about diagnosing why a skill got big:

Non-DRY duplication. Give every part of a skill a single source of truth; don't repeat a piece of reference (a PRD template, a "what is a test seam" note) across multiple steps or places, including across reference material.
Sediment. The classic shared-doc problem: many people contribute to one markdown file, add their own stuff, and don't feel brave enough to delete or modify anyone else's — so irrelevant and stale material accumulates. Fix it by re-examining structure first (is this relevant to all branches? move it to the right branch), then removing what's irrelevant and killing what's stale.
No-ops. Passages that appear to instruct but don't actually change the agent's behavior — especially common when an agent writes the skill. Example: a paragraph telling the agent to "write a long detailed commit message" it would write anyway. The tool is the deletion test: delete the passage; if behavior is unchanged, it was a no-op, so it stays deleted.

The source credits exactly these techniques — deletion tests, compressing into Leading Words, cutting irrelevant material and sediment — for how his skills get "so small." Pruning is the maintenance complement to Minimal Skill Surface's structural smallness.

Claims

Massive skills are usually a symptom of another failure mode (duplication, sediment, no-ops), not a root cause — diagnose the cause rather than just trimming length. observation — reframes "the skill is too big" as a pointer to a specific defect.
Give every part of a skill a single source of truth; don't repeat reference material across steps or places. best practice — context: any non-trivial skill; DRY applied to prose and reference, so a change has one place to land.
"Sediment" — irrelevant/stale accumulation in shared skill docs nobody dares delete — is a distinct bloat source; fix it by re-examining structure, relocating to the right branch, or killing stale content. observation — names a social/organizational bloat mode specific to collaboratively-edited skills.
No-ops are passages that look instructive but don't change behavior (common when an agent authors the skill); find them with a deletion test — remove the passage and check whether behavior changes. principle — durable: the deletion test is a general falsification move — a line earns its place only if removing it degrades the output.
The deletion test is a validation-gated update with n=1 and a human as the score — which means it mechanizes: an optimizer emitting delete edits accepted only on held-out improvement is the same falsification, run at scale. observation — sharpens the deletion test by naming its general form (see Validation-Gated Update, Text-Space Optimization). The substitution is not free: a benchmark score is a narrower falsifier than an author's read of "did behaviour change on the work I actually do."

Validation-Gated Update — the general form of the deletion test: keep a change only if a held-out score strictly improves.
Text-Space Optimization — pruning executed by an optimizer rather than an author.
Minimal Skill Surface — the structural side of smallness; pruning is the recurring cleanup that keeps that surface minimal as edits accumulate.
Leading Words — compressing verbose instruction into a leading word is one of the pruning moves (it removes words while keeping steering).
Agentic Simplicity — "delete anything that isn't making it better" is the simplicity discipline; the deletion test is its concrete falsifier for a skill.
Concise Prompting — the deletion test at the prompt scale: cut redundant procedure a capable model already follows.
Skill Authoring Checklist — pillar 4 (Pruning).
Distillate: Building Great Agent Skills: The Missing Manual
Distillate: microsoft/SkillOpt — training skills like weights, without touching weights — the deletion test generalized: gated delete edits proposed by an optimizer.

Skill Pruning

Claims

Related

Linked from