What problem does it solve? Documents written for AI agents often fail unpredictably: weak descriptions never trigger, bloated files bury key steps, and vague completion criteria let agents stop early. This Skill provides a reference framework for writing any document an agent consumes so the agent follows the same process reliably every run. ## Core Features & Use Cases - Context pointer design: Rules for writing skill descriptions and AGENTS.md lines that front-load trigger words and list distinct branches so agents reach material reliably. - Information hierarchy and progressive disclosure: A ladder for deciding what stays inline versus what moves behind a pointer into disclosed reference files, controlling context load and cognitive load. - Completion criteria and splitting: Guidance on writing checkable, exhaustive step-completion criteria and deciding when to split documents by sequence or invocation, including model-invoked versus user-invoked skills and router skills. - Use Case: When authoring a new Claude Code skill, use this reference to write a description that triggers on the right branches, prune no-op sentences, and decide whether shared reference material belongs in a sibling file like SKILL-MECHANICS.md. ## Quick Start Ask the agent to review your draft skill or CLAUDE.md using the writing-for-agents reference and suggest improvements to its description, structure, and completion criteria.
