Matt Pocock on Agent Skills: How not to fall into skill hell (Category AI4SDLC)

Watch Matt Pocock perform on AI EngineerBuilding Great Agent Skills: The Missing Manual"Skills are discussed as engineering artifacts: with classification, support price and a clear quality criterion. The main idea I took away from it is that good skill is needed to get a predictable process from a probabilistic system. Not the same answer every time, but the same way of working: when to connect, what to read, what steps to take, where to stop, what to consider the completion. This approach is better than the standard view of skills as a warehouse of useful instructions, in which skill hell quickly begins: dozens of files, intersecting rules, old notes, incomprehensible triggers, and as a result, the agent either does not call the desired skill, then drags into context unnecessary, then prematurely declares the work finished.

Pocock's useful frame starts with Invocation is how skill is called. 1️⃣ Model-invoked skill Skill has a description that the model sees in advance and can decide: "Yes, now it is necessary." You have to pay for it with context, because the description always takes place. Therefore, such a skill is justified only if the agent really has to reach him. 2️⃣ User-invoked skill It doesn't hang in context and is called by a person manually. This saves tokens and the attention of the model, but puts the load on the user: you need to remember that such a skill exists. If there are too many such manual skills, a separate router skill appears - not a magic dispatcher, but a human map that helps you choose the right path.

It goes on. internal anatomy skill. In fact, there are two types of material: 1️⃣ Steps actions that the agent must perform in an orderly manner. 2️⃣ Reference Rules, definitions, examples, reference, which the agent reads if necessary.

A good skill does not have to be a procedure. It can be a set of reference rules, like a checklist review. Or vice versa, almost pure sequence of steps. The important thing is that the material should lie at the right level of the information hierarchy. What you need in every launch stays in SKILL. md What is needed only in individual branches, goes to separate files and downloaded through context pointers This is not just token saving, but a way not to smear the attention of the agent.

In the steps, the completion criterion should be defined, so the agent will distinguish “ready” from “nearly ready”. "Take care of the problem" is a bad criterion. “Check any modified models, list any discrepancies found, and indicate which tests cover them.” The murkier the completion boundary, the higher the risk of early completion when the agent finishes the job too hastily.

Matt pays much attention to the use of leading words. (Names of patterns or architectural concepts). These are short anchor words that are already in the pre-training of the model and help stabilize behavior: not a long, repetitive instruction, but a compact term around which the right way of thinking gathers. A good leading word simultaneously helps invocation - the agent more often understands when skill is needed - and execution - the agent inside the skill keeps the same mode of operation.

And finally, Matt talks about how to clean the skills. (pruning)As they degrade as documentation, it is easier to add a new line than to remove the old one. So there are repetitions, cycles and non-working instructions.

The practical conclusion is simple: skill is not a prompt, but a small product interface to the behavior of an agent. It has discoverability costs, support costs, liability limits, and failure modes. It should be designed almost like an API:

• Clear trigger; One source of truth for each rule. Verified completion criteria; Separation of steps and reference; Progressive disclosure for rare branches. Regular cleaning of no-op and outdated instructions.

#AI #AI4SDLC #Engineering #Agents #Software #Management