writing-for-agents

Structure documents with clear steps ending in checkable completion criteria, keep always-loaded pointers tightly trimmed with trigger words first, and disclose branch-specific reference material behind pointers. State target behaviors positively rather than issuing blanket prohibitions.

A context pointer is a reference in the agent's context that names material not currently loaded and encodes the conditions for reaching it. A skill description or an AGENTS.md line naming a document are both pointers, and their wording determines how reliably the agent reaches the target.

Split by sequence when completed steps tempt the agent to rush the current one, or split by invocation when a standalone trigger word should fire on its own. Each new model-invoked skill adds context load through its description, so the independent reach must justify the cost.

Use model invocation when the agent needs to discover and trigger the skill autonomously, keeping a model-facing description. Set disable-model-invocation: true for skills only reached through explicit commands or path pointers, writing the description as a one-line human summary.

Premature completion comes from vague completion criteria that cannot distinguish done from not-done. First tighten the criterion's boundary with checkable, exhaustive wording; only hide later steps across a real context boundary if the boundary stays blurry and rushing persists.

Blanket prohibitions drag the forbidden concept into context without stating what to do instead, and negation is a weak modifier overridden by strongly activated ideas. Prefer stating the positive target behavior, reserving direct bans for hard guardrails or specifically named failure modes.