Define the outcome
State the concrete result, acceptance tests and what “done” means. Avoid endless instructions such as “improve everything”.
Agents & automation
An agent needs a job, boundaries and proof—not just a personality.
Use this blueprint to turn a prompt into a reliable worker that can continue safely.
Limit authority
Name allowed systems, data and actions. Separate read-only diagnosis, reversible changes, destructive actions and spending.
Give evidence rules
Require logs, tests or direct observation for important claims. The agent must distinguish configured, running, tested and user-visible.
Add state and recovery
Document checkpoints, rollback material, retry limits and what happens after interruption.
Control tools
Specify which tools are available, when browsing is required and which credentials must never be copied into outputs.
Set stop conditions
Stop on completion, budget, unsafe uncertainty, missing authority or repeated failure. Escalate with evidence, not vague status.
Start with a tested template