Prompt Optimizer

What is prompt engineering?

Prompt engineering is the practice of crafting input instructions that reliably guide AI language models to produce the output you want. Well-engineered prompts specify a role for the model, describe the task precisely, constrain the output format, and set length expectations — reducing hallucinations and improving consistency across runs.

• Role assignment: Telling the model it's a "senior software engineer" or "professional writer" activates relevant knowledge and style.
• Output format specification: Saying "respond as a markdown table" eliminates ambiguity and makes downstream parsing easier.
• Model-specific techniques: Claude responds well to XML tags; GPT-4 to system/user separation; open-source models to explicit instruction markers.

Why a deterministic prompt optimizer?

Most "AI prompt generators" themselves call an LLM to rewrite your prompt — which introduces variability, API costs, and latency. Our optimizer is fully deterministic: given the same five answers, it always produces the same prompt. This means you can build repeatable workflows, store prompts as artifacts, and version-control your prompt engineering decisions.

The wizard applies a curated rulebook of prompt engineering best practices compiled from published research by Anthropic, OpenAI, Google DeepMind, and the open-source community. No AI calls needed.

Model-specific formatting guide

• Claude: Uses XML tags (<role>, <task>, <guidelines>) for reliable section separation. Ends with "Think carefully before responding."
• GPT-4 / ChatGPT: Separates system and user messages with clear [SYSTEM] and [USER] headers. System message carries the persona and constraints.
• Gemini: Uses flat instruction + task structure with explicit role and instruction lines.
• Open-source (Llama, Mistral): Uses the standard ### System / ### Instruction / ### Response format common to instruction-tuned models.
• Generic: Natural language structure that works across any model.

How to use the wizard

Five steps, most of them one click. Pick your platform, then the task type (code, content, analysis, Q&A, summarization, extraction, or brainstorming), then how you want the output formatted, then how long it should be. The last step is the only one that takes real thought: describe your task in the text box, and use the optional constraints field for rules like tone, language, or things the model must avoid. Hit Generate Prompt and you get the assembled prompt plus a checklist of every optimization that was applied to it. Copy it, or hit Edit constraints to tweak the description without redoing the earlier steps.

Frequently asked questions

Does this send my task description anywhere?

No. The prompt is assembled with plain JavaScript in your browser. Nothing is logged, stored, or sent to an API — which also means you can safely describe internal or confidential work in the task box.

Will the generated prompt work if I paste it into a different model?

Mostly, yes. The structure differs per platform (XML tags for Claude, system/user headers for GPT), but the substance — role, guidelines, task, constraints — is the same. If you switch models often, pick Generic and you'll get a structure that works everywhere.

Why does the output include a role like "You are an expert software engineer"?

Because it works. Assigning a role narrows the model's response style and raises the default quality bar. It's the single cheapest improvement you can make to a prompt, which is why the wizard always includes one based on your task type.

Can I skip the constraints field?

Yes, it's optional. But if you've ever gotten a response in the wrong language, the wrong tone, or with a recommendation you can't use, that's exactly what the constraints block prevents. Two sentences there save a whole retry.