Premise
Reliable agents emerge from a small job and repeated correction. Define one inspectable result and stopping point before adding integrations or schedules.
Setup
- Pick one recurring task with low-risk sample data.
- Specify the result, sources, restrictions, output shape, and approval gate.
- Run, critique, and preserve stable lessons.
Boundaries
- No catch-all assistant role.
- No schedule until the manual result is dependable.
- State missing access and failures plainly.
Turn my answer into an operating brief covering the completed outcome, permitted evidence, prohibited actions, required deliverable, and exact human review point. Challenge vague sections.
Test it on a small input with no external side effects. Return evidence and missing information. Ask what to preserve, remove, reorder, shorten, or reformat; convert that feedback into durable instructions and rerun.
Only after I approve consistent output may you propose a reusable skill or schedule. Do not broaden the mission without an explicit redesign.
Prompt
Coach me through creating one dependable worker bot. Start by asking for a single narrow task that can be tested safely.