Four Dimensions to Evaluate Agent Skills: Matt Pocock's Checklist for Trigger, Structure, Guidance, and Simplification

First Dimension: Trigger — When the Skill Is Invoked

Skills can be invoked in two ways. Model-invoked skills have their descriptions placed in the Agent's context; the Agent decides autonomously whether to load them. Advantage: flexibility and automatic triggering. Disadvantage: each added skill increases context weight, consumes more tokens per request, and the Agent may fail to use a skill when needed or misuse it when not.

User-invoked skills keep descriptions hidden from the Agent; only the user can trigger them manually. Advantage: zero context overhead, no Agent interference. Disadvantage: the user must remember the skill exists and invoke it at the right moment.

Choice depends on goals. For full Agent autonomy with tolerance for token cost and unpredictability, choose model invocation. For a lightweight, controllable Agent, choose user invocation. Matt Pocock prefers user invocation because, despite added cognitive load, it eliminates uncertainty in Agent behavior.

Second Dimension: Structure — Internal Organization of a Skill

A skill consists of two core units: Steps (the execution workflow) and References (supporting information). Some skills have only steps, some only references, most have both.

Example: Matt's 2PRD skill contains three steps — find relevant context, confirm test cases with the user, write the PRD — and two references — test-case说明 and PRD template.

Key technique: move branch-specific content out of the main file. If a skill has multiple usage branches (e.g., does both A and B), place reference material relevant to only one branch into separate Markdown files and reference them via context pointers in the main SKILL.md. This keeps the main file smaller, easier to maintain, and saves tokens.

Third Dimension: Guidance — Making the Agent Follow Intent

Agents often ignore lengthy instructions. The core technique is Nudge Words — terms that pack rich meaning into minimal space. Example: vertical slice in software development means implementing one small complete feature first, then expanding, rather than writing layer by layer. Inserting this term into a skill causes the Agent's reasoning trace to show we will proceed in a vertical slice manner and it actually behaves accordingly.

Advanced technique: hide future goals . If the Agent under-invests in a current step (e.g., in planning mode it rushes to create plan and skimps on clarify questions), split the skill into two independent skills so the Agent sees only one step at a time. Without pressure of a next step, it invests more in the current one. Matt splits create plan into document review (only clarify questions) and 2PRD (create plan). During document review the Agent doesn't know 2PRD exists, so it thoroughly clarifies questions.

Fourth Dimension: Simplification — Removing All Unnecessary Content

Skill entropy is inevitable. Regular cleanup targets three common problems:

Duplication (DRY) : same information appears in multiple places. Keep only one authoritative source per piece of content.

Sedimentation : everyone adds content, no one deletes. Skills fill with might be useful material. Check each section for relevance to all current branches; move irrelevant sections to branch files, delete completely unrelated ones.

No-ops : content that appears to perform an operation but whose removal leaves Agent behavior unchanged. Detection method: deletion test . Remove a section; if the Agent continues doing the same thing, that section is a no-op.

Summary

A good Agent Skill is not measured by lines of Markdown but by correct trigger timing, clear structure, effective guidance, and lean content. Matt Pocock's checklist ties these four dimensions together. Whether writing your own skills or evaluating others', running through this checklist reveals where problems lie.