Why Agent Skills help: reusable steps still need adaptation

Researchers compared no prior experience, cleaned workflow records and SKILL.md distilled from the same records. The experiments cover terminal and tool tasks using a limited set of Codex and Gemini CLI configurations.

The corpus contains 8,135 normalized trials. Taxonomy development retained 238 valid labels from 240 open-coded records, followed by paired-trajectory analysis. The reported 65.7% procedural anchoring and 4.5% knowledge injection describe mechanism analysis, not exhaustive human labeling of all trials. Matched success averaged 6.06 percentage points above Workflow Memory.

With candidate pools growing from 5 to 100, actual-use precision fell from 29.6% to 3.3%. This measures overlap with designated skills, not task success. Retrieval, explicit selection and real execution were independent experiments.

A practical implication is to describe applicability, essential steps and completion checks, then adapt them when the environment changes. This is our interpretation. Rigid commands or assumptions can still fail. The limited terminal-task settings and sampled taxonomy are not proof of universal gains across agents or open-ended collaboration.