Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

Skill Package

Lecture 06-03 · pattern · Reflect × Route 中文 README

Status

pattern.py landed (with lecture 28 / 06-03 of the 极客时间《Agent 设计模式之美》 column). The claim: a skill enters the library only after it passes external verification, and the router only ever routes to VERIFIED skills. Every skill enters as TRIAL — human-written or distilled — and is promoted only when a set of golden questions (deterministic input → expected-output checks) all pass. Post-reuse success rates feed back: a verified skill that starts failing is demoted to TRIAL and must re-verify, which is the staleness guard. Hermes inspires the successful, multi-step-task trigger; this reference interface additionally requires several distinct tools to filter trivial traces. Distillation earns storage, never trust.

Quick start

cd ../payroll-lab
python3 skill_package_lab.py             # scene 1: the gate catches the old policy year; scene 2: the router only sees VERIFIED
python3 skill_package_lab.py --no-gate   # scene 3: stored as trusted without the gate — 209 of 800 bases computed wrong

Where this pattern sits

This pattern sits at the intersection of Reflect (cognitive function) and Route (execution topology). See the two-axis matrix for how it relates to neighboring patterns.

What this pattern covers

The pattern's working title is Skill Package (Chinese: 技能包). Detailed treatment in the Manning book Designing AI Agents (Ch06) and in the column.