A ready-to-use four-part checklist for reviewing skills—trigger, structure, steering, pruning—to help you get out of Skill Hell.
Building Great Agent Skills: The Missing Manual
21 min total·Actually worth watching closely: ~1 min·1 must-watch clip
- 0:00 – 2:00Listen
Skill Hell: why you need a review checklist
The opening lays out the current state: skills are freely available but you can't tell good from bad, and organisations don't know how to turn their operating procedures into things an agent can do. It introduces a four-part shared checklist: trigger, structure, steering, pruning.
Without solving the standards problem of how to judge and write skills, you won't get the bounty that skills can offer.
Purely spoken problem setup and framework overview—no visuals you need to see; works fine as a podcast.▶ Jump to 0:00 - 2:00 – 7:00Skim
Trigger design: user-invoked vs model-invoked
Breaks down the two ways a skill gets invoked and gives the core mental model: the description is a 'pointer' that lives in context, pointing to more context that can be loaded on demand; remove the pointer with disable model invocation: true and the skill becomes user-invoked only.
Choosing user-invoked skills removes the unpredictable-triggering problem outright; keeping model invocation means you need evals to make sure the skill fires at the right time.
This chapter has the talk's only frame flagged as worth viewing (the pointer diagram around 355s); a glance at the slide helps build the mental model, and the rest can just be listened to.▶ Jump to 2:00 - 7:00 – 10:40Listen
Internal structure: steps + reference
Proposes that skills are composed of just two units—steps (the procedure) and reference (supporting information)—and uses the 2PRD skill to demonstrate writing from scratch: define three steps first, then attach two bits of reference.
The order for writing a skill from scratch: work out the steps first, then the reference material each step needs.
Mostly spoken methodology and examples; nothing in this chapter was flagged as depending on visuals.▶ Jump to 7:00 - 10:40 – 14:03Listen
Branch analysis: making the main skill.md as small as possible
Gives a concrete rule for pruning the main file: material used by every branch stays in skill.md, material used by only one branch moves to a separate markdown file inside the skills folder, leaving just a context pointer that says 'if you need this, go to this file.'
Using branch analysis to decide what stays and what moves out is the core technique for controlling the size of skill.md.
Rule-based exposition with clear logic—easy to follow by ear, no demo visuals.▶ Jump to 10:40 - 14:03 – 17:30Listen
Steering: leading words and not enough legwork
Explains that 'the agent isn't doing what I want' is usually down to missing leading words—terms that pack lots of meaning, like vertical slice—whose effect you can verify directly in the reasoning traces; for a step getting rushed through, the fix is splitting it into its own skill and hiding the future goal.
The effect of leading words can be observed and verified in the reasoning traces—the most direct feedback loop for tuning a skill.
Argued mostly through spoken examples like Plan Mode; the value is in the reasoning, not the visuals.▶ Jump to 14:03 - 17:30 – 20:42Listen
Four failure modes and a self-bootstrapping close
Points out that massive skills are a symptom, not the disease: DRY violations, accumulated 'sediment', and no-op paragraphs each have their own remedy (a single source of truth, relocating or killing material, deletion tests); the whole method is finally encoded as the ready-to-use 'Writing Great Skills' skill.
Suspect a paragraph does nothing? Delete it and test—if the behaviour doesn't change, it's a no-op and can be killed for good.
The failure-mode rundown and the close are both spoken summary; just listening gives you the full set of actionable checks.▶ Jump to 17:30