writing-instructions-for-agentslisted

# Writing instructions for agents Apply `writing-for-the-reader`. Here, the reader is an agent choosing and performing behavior from finite context. ## Contract 1. **State the task and trigger.** Make the required behavior and the conditions for using it discoverable in the description and opening lines. 2. **Specify observable conduct.** Describe the action to take, the relevant inputs, and the result or stopping condition. 3. **Keep only behavior-changing context.** Include a pipeline fact when the agent must act differently because of it. Express that difference as direct conduct. 4. **Order by execution.** Put prerequisites before actions, actions before verification, and exceptions beside the step they alter. 5. **Use progressive disclosure.** Keep the governing path compact and point to references for detail that is needed only sometimes. 6. **Preserve signal strength.** Use ordinary declarative language. Let ordering and specificity carry priority. ## Maintenance tests - **Contribution:** compare behavior with and without the instruction. Keep it when behavior changes usefully. - **Environment:** prefer a query or tool over prose that can become stale. - **Duplication:** give each instruction one canonical home and use a pointer elsewhere. - **Generalization:** state the underlying behavior rather than the session that revealed it. - **Control boundary:** compare the reader's inputs, action, validation, and output with each context sentence