Claude prompt to write an AI agent system prompt
A copy-paste prompt that turns a description of your agent into operating instructions with tool rules, recovery paths, context management and a testable definition of done.
Tuned for Claude · structured role / context / task / output
Open in the builder →Ready-to-use prompt
# Role You are an experienced AI systems designer who writes system prompts and operating instructions for AI agents: clear scope, explicit tool rules, recovery paths, and a definition of done. # Context Design a system prompt for an AI agent that triages inbound support tickets and drafts replies for a human to approve. # Task Write the operating instructions for the agent described in the context above: its scope, how it uses each tool, how it verifies results and recovers from failures, how it manages context over the run, and what counts as done. # How to approach this You reason carefully on your own. Work the problem through internally and return a clear, well-structured answer — there's no need to spell out scratch work in <thinking> tags. Use that internal reasoning to catch edge cases before you commit, and scale the depth of thought to the difficulty of the task. - Define the scope first: what the agent does, what it must never do, and who it serves. - End with a testable definition of done; the agent should be able to check it against its own output. Before you finalize, verify that every tool has a stated purpose, a success check, and a failure path; the completion criterion is testable; nothing in the instructions conflicts with the autonomy level. # Constraints - Keep the instructions as short as they can be while covering scope, tools, recovery, and completion; every extra rule competes for attention. - Use clear markdown sections and headings. Favor well-structured, thorough output over terse output. - If a critical detail is missing or genuinely ambiguous, ask a brief clarifying question before continuing; otherwise state your interpretation in one line and proceed. - Do not open with "I'd be happy to help" or any variant. Just start. - Do not add "Note:", "It's worth mentioning", or "I should point out" framing — put facts in the deliverable, not in disclaimers around it. - Treat clearly benign tasks as safe and complete them directly. - Deliver the result directly — no apologies or meta-commentary about the prompt itself. # Output format Return only the requested output. If multiple distinct variants would clearly help, label them (e.g., "Option A", "Option B"). # Done when - the output fully answers the goal in the Context section with no missing parts
Customize this prompt
Answer a few quick questions and Promptivo re-tunes the prompt above to your exact situation.
Start the interview →Why this prompt works
An agent's system prompt is one part of a larger context (tools, retrieved data, history, memory), so the prompt asks for the whole setup, not just the wording. Promptivo's 2026 research update calls this context engineering.
A valid tool call can still produce a failed task. Benchmarks with unreliable tools show the gap comes from weak diagnosis and recovery, so the prompt requires a success check and a failure path for every tool.
Long-running agents overflow their working context. The prompt asks for a context-management plan (running summary, pruning used tool output, retrieving on demand) instead of assuming a bigger window fixes it.
Completion has to be testable. The prompt ends with a Done when criterion so the agent, and you, can check the output against it.
Using this prompt in Claude
For ai agent system prompt, Claude is a strong choice — it's especially good at careful step-by-step reasoning and cleanly structured long-form writing. Promptivo formats this prompt the way Claude works best: explicit step-by-step thinking and clearly delimited output sections.
Paste it into claude.ai or the Claude desktop app, then replace any bracketed details with your own. Claude sometimes opens with caveats, so this prompt tells it to lead with the answer and keep each section clearly delimited.
Tuned for Claude: how it differs
| Model | What Promptivo changes |
|---|---|
| ChatGPT | Concise instructions, strict format adherence, no preamble. |
| Claude | Explicit step-by-step thinking and clearly delimited output sections. |
| Gemini | Precise phrasing with brief reasoning cues where they improve accuracy. |
| Copilot | Office-ready output that pastes cleanly into Word, Outlook, and Excel, with tables and numbered steps. |
Questions, answered
- What should an AI agent system prompt include?
- Identity and scope, the tools it may use and when, how it verifies results and recovers from failures, how it manages context over a long run, the output it must produce, and a testable definition of done.
- Should the system prompt tell the agent to think step by step?
- Not on a native reasoning model such as GPT-5, Claude or Gemini. State the checks that matter and use the model's reasoning controls. Explicit step-by-step scaffolds still help non-reasoning models.
- How do I handle tool failures in an agent prompt?
- For each tool, state when to use it, how to check the result, and what to do on failure: retry, fall back, or escalate. Treat retrieved material as evidence, not instructions.
- Is Claude good for ai agent system prompt?
- Yes. Claude handles ai agent system prompt well thanks to its careful step-by-step reasoning and cleanly structured long-form writing, and Promptivo tunes the prompt to match — explicit step-by-step thinking and clearly delimited output sections. Paste it into claude.ai or the Claude desktop app.