Dayward AIWeChat Mini Program
Agent Skills in 7 Days: Turn Experience Into Reusable Capability
Week 1 · D3🔒 LockedAbout 5 hours

A Design Method: Distilling From Repeated Tasks, Checklist Style vs. Reference-Manual Style, Four Anti-Patterns, and Trigger Testing

Turn writing a skill from inspiration into a method: judge whether a task is worth turning into a skill, organize content into one of two mature forms, avoid four anti-patterns, and use a set of positive and negative examples to measure description's trigger rate.

Today's goals 0/3

Sign in to tick these off and save your progress.

Today's Goals

  1. Judge whether a repeated task is worth turning into a skill, using three criteria
  2. Distinguish checklist style from reference-manual style, and pick the right form for a task
  3. Design a set of positive and negative example queries, measure description's trigger rate, and rewrite it accordingly

Yesterday you wrote your first skill by trying three sentences and seeing whether it triggered. That works with one skill and stops working by the tenth — you cannot hold in your head whether ten descriptions compete with each other, and you cannot tell after editing one sentence whether you improved it or broke it. Today gives you a ruler you can reuse. Once you have read the walkthrough and finished the lab, scroll back to the top and tick off the three goals.

Plain-Language Walkthrough

Three criteria for choosing a subject

The filing cabinet's space is finite — not physical space but the model's attention. Every extra skill adds a line to the catalog and raises the chance of a spurious trigger. So the first question is never "how do I write this skill" but "is this thing worth turning into a skill at all."

Three criteria, and all three must hold before you start:

First, repetition. You have done this at least three times and will do it again. Something done once has no experience in it — you do not even know whether you got that once right. The test is plain: can you say immediately when you last did this? If you cannot, it is not as frequent as you thought.

Second, a correction exists. The first time the model did this, you interrupted it. This is the hardest of the three, because it proves two things at once: the model genuinely does not know (or you would not have corrected it) and you genuinely do (or you could not have). A skill with no record of correction will most likely come out as a pile of correct-and-useless lines — "handle errors appropriately," "follow best practices," sentences the model would produce without your skill anyway.

Third, the result is checkable. When it is done you can tell whether it was done right. Whether a commit message matches the format is visible at a glance; whether the amounts on a filled-in report are correct takes one calculation. Conversely, "think of a better architecture for me" is a task whose correctness is known three months later — not that it cannot be a skill, but you cannot iterate on it, so once written you can only feel that it is useful.

Start only once all three pass. Fail any one of them and the time you spend on the skill will most likely be wasted: what is not repeated goes unused, what has no correction comes out as filler, and what cannot be checked leaves you never knowing whether it worked.

These three criteria filter out far more than they keep. That is correct — three folders in the cabinet consulted daily beat thirty nobody opens. Which raises the next question: for the thing that passed, how should the content be organized? As a procedure followed step by step, or as a manual consulted at will? Those two forms suit entirely different situations, and picking wrong gives you something usable but awkward. Below, both forms and the four most commonly hit anti-patterns, settled in one go.

Sign in and link WeChat to read

Sign in, then link your WeChat mini program account on the account page. Linked accounts share one free article per day. Unlock additional articles with rewarded videos in the mini program.

Mini program QR code

Scan in WeChat to open Dayward AI. Open the same course and chapter, choose the rewarded video, then refresh this page after completion.

Interview questions

  • Which tasks are worth turning into a skill and which are not? Give me a test I can apply on the spot.什么样的任务适合做成 skill,什么样的不适合?给我一套能当场用的判断标准。
    Common in ChinaCommon overseasBasic#agent-skills#skill-design

    🔒 Unlock this chapter to see the answer keys.

  • How do you evaluate whether a skill description is good? Is trying a few prompts yourself enough?怎么测一个 skill 的 description 好不好?自己试几句话够吗?
    Common in ChinaCommon overseasIntermediate#agent-skills#evaluation

    🔒 Unlock this chapter to see the answer keys.

  • When optimizing a description, how do you avoid overfitting to the very queries you wrote?优化 description 的时候怎么避免过拟合到你自己写的那几条测试查询?
    Common in ChinaCommon overseasDeep dive#agent-skills#evaluation

    🔒 Unlock this chapter to see the answer keys.

Comments