A useful skill checklist starts with how the skill is invoked, because model-invoked and user-invoked skills move cost between the agent and the user.
2
Skills should separate procedures from supporting reference material, then keep branching references outside the main skill.md file behind context pointers.
3
Agents follow skills more reliably when they use dense leading words, focus on one phase at a time, and lose duplicated, stale, or behaviorally irrelevant instructions.
Summary
Matt Pocock presents a checklist for escaping what he calls "skill hell", the problem of having many available agent skills without knowing which ones work or how to maintain them. The checklist covers trigger, structure, steering, and pruning. Trigger design means choosing between model invocation, which adds context load and unpredictability, and user invocation, which gives the pilot more control but requires more knowledge. Structure divides a skill into steps and reference material, with branch-specific references moved out of skill.md. Steering uses repeated terms such as "vertical slice" to shape the agent's behavior. When a step needs more investigation, splitting it into a separate skill can hide the later goal and make the agent do more work first. Pruning removes duplication, accumulated sediment, stale material, and no-ops. Pocock has encoded the framework in his "writing great skills" skill and points viewers to his repository.
Skill hell comes from having no shared way to judge skills
Pocock describes "skill hell" as having many downloadable and reusable skills without understanding how they fit together or how to tell a good skill from a bad one. People try several frameworks at once and fail to get the results promised by the skills. Organizations face the same problem when they try to turn operating procedures into actions an agent can perform. He says the missing piece is a shared rubric. The talk supplies a checklist for examining whether a skill does what it claims and for improving it.
Invocation moves cost between the agent and the pilot
A skill can be user-invoked manually, or model-invoked through a description that remains in the agent's context. Model invocation gives the agent more freedom, but every available description adds context load and another choice for the agent. Pocock says a hundred model-invoked skills would mean a hundred descriptions in context. User invocation reduces that load and avoids the model deciding not to follow a pointer, but it increases the pilot's cognitive load because the user must know which skills to call. Neither choice is free.
A skill becomes easier to design when it has steps and reference
Pocock divides most skills into two units: steps, which describe the procedure, and reference, which contains supporting information. His 2 PRD skill finds relevant context, confirms test seams with the user, and writes a product requirements document. Its supporting material explains test seams and contains a PRD template. A skill can contain only steps or only reference, but treating these as separate units makes the design easier to inspect and revise.
The main skill.md file should contain only material needed across branches
Pocock argues that a small skill.md file is easier to maintain and audit, while also reducing the tokens used on every request. Branch-specific material should move behind context pointers into separate files. His domain modeling skill can update context.md, create architectural decision records, or do neither, so the corresponding templates do not need to sit in the main file. The pointer tells the agent where to find a template when that branch requires it.
Leading words give agents a compact instruction for a pattern of behavior
Pocock's main steering technique is to use leading words, terms that carry a lot of meaning in a small amount of text. He uses "vertical slice" as an example. Agents often implement a large task layer by layer, starting with the database and moving through schemas, APIs, and the front end. Repeating "vertical slice" can invoke the development concept of building a thin working slice across the system. Pocock says the technique can be checked in reasoning traces when the agent repeats the term and produces plans shaped around it.
Separating phases can make an agent do more work before reaching its goal
Pocock says agents often do too little work in an early step when they can already see the final objective. In plan mode, an agent may ask only a few clarifying questions before eagerly writing a plan. His solution is to put the clarification phase in a separate skill called grill with docs, then run 2 PRD afterward. The agent sees one phase at a time, so the future goal is hidden while it investigates the current phase. He says this is especially useful when a particular step needs more effort.
Pruning requires deletion, single sources of truth, and branch cleanup
Pocock identifies several causes of oversized skills. Repeated instructions and reference material should have a single source of truth. Sediment builds up when people keep adding to a shared markdown file without deleting old material, so irrelevant or stale content should be removed or moved into the right branch. No-ops are instructions that look useful but do not change the agent's behavior. His deletion test asks what would happen if a paragraph disappeared. If the agent would behave the same way, the paragraph can go.
The full checklist checks invocation, layout, behavior, and excess
The final sweep starts by checking whether the skill fires at the right times and whether its design creates context load or cognitive load. Structure means separating steps from reference and moving branch-specific material outside skill.md. Steering means using consistent leading words and checking whether they appear in reasoning traces, then deciding whether phases should be split to increase legwork. Pruning removes sediment, crud, duplication, and no-ops. Pocock directs viewers to the writing great skills skill as a practical way to apply the checklist.
"The cool thing about this technique is you can know if it's worked because you say vertical slice in your skill, and then you'll notice in the reasoning traces that it's saying, "Okay, we're going to do this as a thin vertical slice.""13:57
Who should watch
You are building agent skills and need a way to decide whether instructions belong in the main file or in supporting references.
Your agent has many available skills and you are weighing automatic invocation against keeping control in the user's hands.
Your skills have grown through repeated edits, or your agent rushes through investigation before it reaches a plan or implementation.
