writing-for-agents

Front-load the leading word that does the triggering work, list one trigger per distinct branch the document handles, and cut identity the body already carries. Every word of an always-loaded description costs context on every turn, so prune it harder than the body.

A model-invoked skill keeps a description so the agent can fire it autonomously, paying permanent context load for discoverability. A user-invoked skill sets disable-model-invocation: true, removing the description from the agent's reach so only a human typing its name can invoke it.

Split by sequence when visible later steps tempt the agent to rush the current one, and split by invocation when a distinct trigger word should fire a skill on its own or another skill must reach it. Each split spends context or cognitive load, so the cut must earn it.

Premature completion happens when a step's completion criterion is vague and later steps remain visible, pulling attention toward being done. Sharpen the criterion's bound first; only hide later steps across a real context boundary if the rush persists.

Steering by prohibition drags the forbidden behavior into context and makes it more available, since the negation is a weak modifier the activated concept overruns. State the positive target behavior instead, reserving prohibitions for hard guardrails paired with the positive form.