# Prompt Engineer: Precise Prompts With Clear Success Criteria > The Prompt Engineer is an AI prompt engineer in the Ai1 platform by MyZone AI: it turns a rough brief into a prompt tuned to the text, image or video model you use, with written success criteria, a never-permit list and test cases, and it writes prompts rather than running them. Canonical page: https://myzone.ai/pages/agents/ai-prompt-engineer-agent Part of Ai1, by MyZone AI. Book a Prompt Engineer walkthrough: https://calendly.com/d/ct6h-tcy-8qf/ai1-demo?a1=Prompt%20Engineer&utm_source=myzone.ai&utm_medium=agent-page&utm_content=ai-prompt-engineer-agent-final Last updated: 2026-10-05. Reviewed by the MyZone AI team. Prompts tuned to your model, with rules for what they must never do. ## At a glance - **You tell it:** What you need the AI to do, as roughly or as fully as you like - **Models it writes for:** Text agents, image generators and video tools - **With every prompt:** Success criteria, a never-permit list and test cases - **Measured scores:** Run by the Evaluator against the test cases it writes - **Your prompts:** Kept in a versioned library with each one's model, settings and history - **What it never does:** Change your live assistants or tools; a person on your team makes the change ## What the Prompt Engineer does - **Writes prompts for your specific model:** Takes your brief, asks targeted questions only when something important is missing, and writes a prompt shaped for the model you use rather than a generic template. - **Sets out what success looks like:** Every prompt comes with examples of good output, a list of things it must never do and test cases that can be used to check it. - **Repairs prompts that underperform:** Works out why an existing prompt gives poor results, rewrites it and explains what changed and why. - **Builds safety into the wording:** For customer-facing prompts or ones linked to live tools, it writes the guardrails and never-permit rules straight into the prompt. - **Moves prompts to a new model:** Ports your prompts when you switch models, removes what will not carry over and asks the Evaluator for a regression check where quality matters. - **Keeps a prompt library:** Organises your prompts into a versioned collection, each with its model, settings and change history. ## How the Prompt Engineer differs from a prompt template A template gives you wording and leaves you to find out whether it works. The Prompt Engineer asks targeted questions when a brief is unclear, recommends a model and settings, and writes a prompt shaped for that model. Every prompt comes with success criteria, a never-permit list and test cases, so the Evaluator can measure it. It writes prompts only, and a person on your team makes any live change. How this differs from the Evaluator (https://myzone.ai/pages/agents/ai-evaluator-agent): the Evaluator runs the measured tests and scores, while the Prompt Engineer writes the prompt, its success criteria and the test cases the Evaluator runs. ## How it works 1. **Share your brief** (Whenever you need a prompt): Tell it what you need the AI to do, as roughly or as fully as you like. 2. **Clarify if needed** (you approve) (Only if something is unclear): If something material is unclear, it asks up to five targeted questions and you confirm the answers before it writes anything. 3. **Choose the right model** (After you confirm): It recommends the model and settings that suit the task, and explains why. 4. **Write the prompt** (Once the model is chosen): It drafts the prompt tuned to that model, with the right structure and examples. 5. **Define success and safety** (With every prompt): It writes down what good output looks like, what the prompt must never do and the test cases to check it. 6. **Deliver, test if needed** (When the prompt is ready): You receive the prompt and its criteria. When you want a measured score, the Evaluator runs the tests. ## When to use it - **Your chat assistant says things it should not:** It rewrites the prompt so the assistant stays on brand and never offers something like a refund without a person's approval. - **An image prompt keeps producing the wrong style:** It diagnoses the problem, rewrites the prompt and notes what changed so you can see the difference. - **You are moving to a different AI model:** It ports your main prompts, versions them and asks for a regression check where the output matters. - **Your prompts are scattered across documents and chats:** It gathers them into one versioned library you can browse and reuse. - **You are not sure which model suits a task:** It recommends a model and settings and explains the trade-offs in plain terms. ## What you get - A finished prompt tuned to your chosen model, ready to copy and use - Written success criteria and a never-permit list for each prompt - Test cases ready for the Evaluator to run - A note on what changed and why when a prompt is rewritten - A versioned prompt library with each prompt's model, settings and history ## Example: support assistant prompt rewrite Example with a fictional company. Names, people and figures are invented to show the agent's output. Any resemblance to a real company or person is unintended. Industry facts are real and cited with their sources. What it was asked: Our chat assistant keeps promising refunds and guessing delivery dates: fix the prompt, keep it sounding like us, and make sure it never offers a refund without a person's approval. The Prompt Engineer read the shop's current prompt and 20 example chats the shop shared (names and contact details removed first), asked two questions, then rewrote the prompt for Claude with success criteria, a never-permit list and 12 test cases, citing the model provider's own guidance. It did not run the prompt or score it: the measured test run belongs to the evaluation agent, and the support lead approves and a person on the shop's team makes the change to the live assistant. Run date: 30 September 2026. ### What it found - A "do whatever it takes to keep the customer happy" line caused the refunds: 6 of the 20 shared chats offered one. Refunds now always go to a person. - Delivery dates were guesses because nothing told the assistant to use the order lookup. The new prompt quotes dates only from the lookup and says plainly when it has none. - Fourteen scattered tone rules, two contradicting, became five rules and three example replies. Twelve test cases are written for the evaluation agent, which runs the measured score before launch. ### Sources - [Prompting best practices - Claude Platform Docs](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices), Claude Platform Docs (accessed 2026-09-30) - [Reduce hallucinations - Claude Platform Docs](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations), Claude Platform Docs (accessed 2026-09-30) ## Guardrails - It writes prompts and criteria only. It does not run the prompt or score it; measured testing is the Evaluator's job. - It does not change your live assistants or tools. A person on your team approves the new prompt and makes the change. - It asks targeted questions when a brief is unclear instead of guessing. - Changes to how the Prompt Engineer itself works go through an approved improvement process, with a record of each change. ## Frequently asked questions ### What are prompt success criteria? The written test of a good prompt: examples of the output you want, a list of what the prompt must never do and test cases to check it against. The Prompt Engineer in Ai1 by MyZone AI writes them with every prompt. The Evaluator then runs the prompt against those test cases and reports a measured result, so you know a new prompt actually performs. ### Can prompt engineering fix an image prompt that keeps missing the style? Yes. The Prompt Engineer works out why the prompt gives poor results, rewrites it for the image model you use and notes what changed and why. It covers text agents, image generators and video tools, using the right approach for each medium. It writes the prompt only and hands it to you or to the tool you use; your image tool creates the picture. ### How does prompt engineering stop a chat assistant from overpromising? By writing guardrails and never-permit rules into the prompt itself, such as never offering a refund without a person's approval. The Prompt Engineer adds them to customer-facing prompts, and a person on your team approves the new prompt and makes the change to the live assistant. ### What does an AI prompt engineer give me that a template does not? A prompt shaped for the exact model and task you have, plus written success criteria, a never-permit list and test cases. A template gives you wording; good prompt engineering gives you a prompt you can check. When an AI provider updates its models, it can check the provider's latest guidance, flag what has changed and update your prompts. ## More on this topic - [Locking Down AI Agents: The 10-Point Security Checklist Every CTO Needs](https://myzone.ai/pages/blog/ai-agent-security): AI agents are powerful,and vulnerable. Learn the 10 critical security steps to lock down your AI agents before they become your biggest liability. ## About Ai1 Ai1 is the AI operations platform by MyZone AI, where each client runs on its own private server. The Prompt Engineer is included on every Ai1 level, including Developer Core, with no per-agent charge. It works alongside the other Ai1 agents on your account. Pricing: https://myzone.ai/pages/services/ai1-pricing. Security: https://myzone.ai/pages/security.