writing-for-agents

Change agent behavior with minimum context in skills, AGENTS.md, CLAUDE.md, and references. For skills, read SKILL-MECHANICS.md.

Depth

A context pointer names external material and when to read it; it spends context load. Progressive disclosure uses the shallowest depth:

1. Inline universal steps.
2. Inline reference needed during them.
3. Link branch-only detail with a precise trigger.
4. Treat the environment as a source of truth; cache only expensive lookups.

Front-load pointer triggers, name each branch once, and never hide mandatory steps in references.

Executable instructions