Skills are context compression
Agent systems · Context architecture
What you will leave with
Extract one repeated protocol, write its trigger and procedure, and verify workers invoke it consistently.
What you will build
A compact catalogue entry backed by an executable, versioned procedure and fixtures.
- 01MineFind repeated protocol and real failure history.
- 02BoundState when the skill applies and when it does not.
- 03WriteEncode ordered actions, proof, stops, and escalation.
- 04InvokeReference by name rather than copying the body.
- 05TestExercise trigger, happy path, refusal, and failure output.
Build it in sequence
Extract a protocol that already works
Start with repeated instructions you have run successfully and corrected after failure. Do not use the skill file to invent an untested operating model.
Write a discriminating description
The description is routing logic. Name the exact task shape and outcome; include exclusions when adjacent skills might trigger. Vague descriptions either load too often or disappear when needed.
---
name: isolated-lane
description: Use for an independent code slice that owns disjoint paths and must return a verified PR.
---
1. Read the ownership contract before opening files.
2. Stop if the task requires a path outside `owns`.
3. Work in the assigned worktree and branch.
4. Return commit, checks, artifact, and any decision request.Encode proof and refusal
A procedure needs required evidence, stop conditions, and the common rationalizations that tempt workers to skip them. “Tests probably unrelated” is a failure mode, not an exception.
Remove the pasted duplicate
Change dispatches to invoke the skill by name and keep task-specific scope in the task. Leaving both copies preserves drift and pays the context bill twice.
Test routing and execution separately
Use prompts that should and should not trigger the skill, then fixtures that force a red gate, out-of-lane request, and unavailable verifier. Measure invocation and protocol completion.
Definition of done
- The skill encodes a proven repeatable procedure.
- Its description distinguishes relevant and irrelevant tasks.
- Steps include evidence, stops, and escalation.
- Dispatches reference the skill instead of duplicating it.
- Trigger and failure fixtures are versioned.
What breaks and what it means
The skill loads for almost every task.
Narrow the description and split unrelated responsibilities.
Workers invoke it but still skip verification.
Make evidence and stop conditions explicit, then add a failure fixture that exposes the shortcut.
Questions, corrections, and useful disagreement
Sign in with GitHub to join the conversation. Comments are public. Article views are anonymous and counted once per browser, per article, each day.