1. Watch it guess
Run the agent without the context. It uses the wrong convention or makes an assumption.
Overview
By the end of this act you can: make the agent follow your project’s conventions and pull in reusable workflows on demand. You’ll have watched it guess wrong without that context, then get it right once you supply it.
Act 0 showed an agent with no idea what your project values. This act fixes that. Both primitives here answer one question. What does the agent know? And both earn their keep by loading only what’s relevant, when it’s relevant.
- Steering files are durable knowledge about this project: conventions, standards, structure.
- Skills are sets of knowledge and reusable workflows the agent pulls in on demand, based on what you’re asking for.
Where the agent’s knowledge comes from
Section titled “Where the agent’s knowledge comes from”Those two are layers in a bigger stack. Picture a pyramid of where the agent gets what it knows, from the always-on foundation at the bottom to the just-in-time specifics at the top:
Skills
pulled in on demand, only when your request needs them
pulled in on demand, only when your request needs them
Steering
your project’s conventions, loaded always or by file
your project’s conventions, loaded always or by file
AGENTS.md
the always-on foundation the agent reads first
the always-on foundation the agent reads first
The lower a layer sits, the broader it is and the more often it loads. The higher you go, the more
specific it gets, and the longer it waits until it’s actually needed. This act is about the two
layers you shape by hand, steering and skills. The base, AGENTS.md, is the project-wide
foundation the agent reads every session.
The method: feel the absence, then the fix
Section titled “The method: feel the absence, then the fix”2. Add the context
You add the steering file or skill and re-run. The behaviour changes in front of you.