Loading Skills: Scanning, Progressive Disclosure, and Trigger Judgment — Bringing Experience to the Table on Demand
Turn reusable experience into skill packs: implement a skill loader that reads only a summary at startup, reads the full body only when a trigger condition matches, and runs an attached script only when needed — and use real numbers to show exactly how much context this progressive disclosure saves.
Today's Goals
- Implement a skill scanner and metadata parser, and explain the directory convention and validation points
- Implement the three stages of progressive disclosure, and quantify its benefit in token count
- Design trigger judgment that avoids a skill firing too often or not firing when it should
Yesterday you plugged in someone else's tools. Today you plug in someone else's experience. When you finish, scroll back up and tick the three goals off.
Plain-Language Walkthrough
A craft card the old hand wrote, read only when needed
Every workshop has an old hand carrying rules nobody else knows: this machine idles two minutes after you switch it on, that material needs a slow feed. He is retiring, so he writes them on cards, one job per card, pinned to the wall. The new hire does not memorize the stack, and should not — it is enough that when he hits that particular job he knows there is a card about it, walks over and reads it.
For fifteen days we gave mca capability: read files, edit files, run commands, use other people's tools. Today we give it experience: why division in this repo must throw when the divisor is zero, how the tests are run here, which step comes first before a release. A model cannot know these. They are not general knowledge; they are what your team settled on after an argument one afternoon.
So why not paste them into the system prompt? Because the wall fills up. Thirty craft cards is normal for one team, a thousand words each, thirty thousand words carried on every single request — even when the user only asked what a function does. Money is one half of it. Attention is the worse half: with thirty unrelated cards spread in front of it, the model is measurably less sensitive to the one that matters.
So the real problem today is not "how do I load a document" but how do I load it only when needed, while still letting the model know the card exists. Those two look contradictory — how do you know it exists without loading it — and resolving that is exactly what progressive disclosure is for.
Interview questions
How many stages does progressive disclosure for skills have, and what triggers each one?技能的渐进披露具体分几个阶段?每个阶段的判据是什么?
Common in ChinaCommon overseasBasic#progressive-disclosure#skills🔒 Unlock this chapter to see the answer keys.
Where is the line between a skill and a tool, and how do you decide which one a given capability should be?技能和工具的边界在哪?同一件事你怎么决定做成技能还是做成工具?
Common in ChinaCommon overseasIntermediate#skills-vs-tools#agent-design🔒 Unlock this chapter to see the answer keys.
How would you design skill trigger conditions, and which is harder to diagnose - over-triggering or under-triggering?技能的触发条件怎么设计?过度触发和漏触发哪个更难查?
Common in ChinaCommon overseasDeep dive#trigger-design#false-positives🔒 Unlock this chapter to see the answer keys.
