All categories/Agents and automation
Level Advanced · System prompt Agent Bounds
Write a system prompt: role, bounds, output, refusal
Your agent prompt reads like a wish list, so it answers everything and nails nothing.
Prompt
Write a system prompt for an agent.
Its job: {{duty}}
Who uses it: {{users}}
Tools it has: {{tools}}
Hard prohibitions: {{forbidden}}
Output format: {{format}}
Five sections, each with explicit criteria:
1. Role and goal: one sentence on why it exists and what success means.
2. Workflow: the order it handles a task in, and which step must stop and ask a human.
3. Tool use: which tool when, and when a tool must not be used.
4. Output contract: format, length, required fields, banned phrasing.
5. Bounds and refusal: which request classes it refuses, with the exact words to use.
End with 5 test cases that show whether the prompt actually works.What to state before it starts
Leave these unsaid and the model fills them with an assumption. Most 'wrong answers' start here.
- {{duty}}: write 'turn X into Y', not 'help users'
- {{forbidden}}: be concrete, e.g. 'never invent a price'
- {{format}}: field names, order, length caps — the tighter the better
Go through each of these
Where this kind of work most often goes wrong. Already in the prompt, so you do not have to remember it.
- 'Which step must stop and ask' is the line between reliable and reckless.
- Refusals need exact wording, or it improvises something vague and polite.