Agents

AI Agent Prompt Builder

Describe the job you want an AI agent to handle and get a system prompt in five labelled blocks: persona, objective, process, restrictions, and what to do when it is unsure. Keeping identity and hard rules apart is the point, so you can tighten a restriction without rewriting the character. Paste the result into a ChatGPT Custom GPT, a Claude Project, or any agent platform.

  • Free to use
  • No sign-up
  • Instant results
0
Agent system prompt
Your result will appear here Add your input, then build prompt

How it works

  1. 01 Describe the task the agent should handle.
  2. 02 Pick the closest agent type, or Custom.
  3. 03 Generate the prompt, delivered as five separate labelled blocks.
  4. 04 Edit any single block without touching the others, then paste it into your agent platform.

What makes an agent prompt hold up

Most system prompts fail the same way: they describe a mood instead of a job, and their rules cannot be checked.

Do

  • Give the persona a specific working identity, not a superlative — "a billing lead who has answered the same refund questions for six years"
  • Write every restriction so you could tell from a transcript whether it was obeyed
  • Start restrictions with Never, Only, or Always
  • Name the exact fallback: what it says, what it asks for, who it escalates to
  • Keep one objective; fold the rest into the process steps

Don't

  • Don't write "be accurate and professional" — it cannot be checked or enforced
  • Don't let the fallback be "use your best judgement"
  • Don't mix rules into the persona block; you will lose track of both
  • Don't pad it. A 900-word prompt buries the four rules that actually matter

Frequently asked questions

Because they change for different reasons. Persona is who the agent is and how it speaks; restrictions are the hard limits it can be held to. When they are mixed into one paragraph, tightening a rule means rewriting the character, and it gets hard to tell whether the agent broke a rule or was just being itself. Split, you can edit either one alone.

PERSONA (who it is and how it speaks), OBJECTIVE (the single outcome it owns), PROCESS (the ordered steps it takes), RESTRICTIONS (hard limits, each one checkable from a transcript), and WHEN UNSURE (its exact fallback behaviour).

It is held to roughly 200-320 words on purpose. Long system prompts dilute attention and bury the rules that matter, and every line that would leave the agent behaving identically if deleted is a line making the important ones harder to follow.

No. It writes the system prompt that defines how an agent should behave. You paste that into a platform such as a ChatGPT Custom GPT, a Claude Project, or your own code to actually run it.

No. Describe the task in plain English and you get instructions you can paste into a no-code agent platform. If you want runnable code instead, use the Prompt-to-Agent Scaffold.

Those mark real details the prompt needs that you did not supply, such as a product name or an escalation contact. They are left blank rather than invented so nothing false ends up in your agent's instructions.

Related tools

Engineering Prompt-to-Agent Scaffold Describe an agent and get runnable starter code, or a skill prompt for your AI IDE. New Agents Document Chat Ask questions about a document and get answers only from its text. New Engineering AI Code Reviewer A real diff viewer that flags issues on the exact lines you changed.