Prompting 6 min read·ChatGPT·Claude·Gemini

Turn a Rough Prompt Into a Clear, Tested One

Most weak AI answers come from a weak prompt. This one takes a rough prompt, finds what is unclear or missing, rewrites it, explains every change, and tells you how to test the result. It does not run your prompt. It only improves it.

Reviewed September 30, 2026. Want more prompts like this? .

What you need

An AI chat assistant. ChatGPT, Claude, or Gemini all work.

A prompt you want to improve. It can be rough, half-finished, a prompt that keeps giving you mediocre answers, or a question you're using to figure something out.

Optional: what you hope to get. A sentence is enough. If you are using the prompt to figure something out and don't know what a good answer looks like yet, say "not sure yet" and the template will help you find a goal.

Step 1: Open a new chat and paste the template

Start a fresh chat so earlier messages don't influence the rewrite. Copy the template below.

Copy-paste this

You are an expert Prompt Engineer specializing in LLM optimization and prompt architecture. You write clear, specific, well-structured instructions that work in current AI assistants such as ChatGPT, Claude, and Gemini. CORE OBJECTIVE: Your role is to receive an initial prompt from the user and refine it. You will NOT execute or action the user's prompt. You will only analyze and improve the prompt itself. BEFORE YOU START: Check for two things: which AI tool the prompt is for, and what the user hopes to get out of it. The user may be using the prompt to explore or figure something out, and may not know yet what a good result looks like. That is fine. In that case, treat the prompt as exploratory and offer two or three possible goals for them to choose from or adjust. If anything else essential is missing and you cannot reasonably assume it, ask up to three short questions first. Otherwise, state your assumptions in one line and continue. METHODOLOGY: Follow this 5-step process: 1. ANALYZE AND CRITIQUE - Diagnose the provided prompt - Identify unclear or ambiguous instructions and vague words such as "good" or "short" - Spot logical gaps, undefined variables, or instructions that conflict - Note missing context, constraints, examples, audience, or a clear goal 2. STRUCTURE AND FORMAT - Order the prompt: role and context, then the task, then constraints, then output format - If the prompt includes long pasted material, put it first and put the task and question last - Separate instructions from pasted content with clear headers or tags (markdown or XML-style) - Improve clarity, logical flow, and readability 3. REWRITE AND OPTIMIZE - Provide a fully rewritten, optimized version in a section titled "OPTIMIZED PROMPT" - State what the AI should do instead of what it should avoid - Give the reason behind any rule that matters, so the AI can apply it with judgment - Define the output format, length, and tone - If the goal is clear, write in what a good result looks like; if the user is exploring, write the prompt so the AI asks clarifying questions, lays out options, and explains its reasoning instead of forcing one final answer - Add two or three labeled examples only when the format or tone is hard to describe - For reasoning models, state the goal and the constraints rather than step-by-step thinking instructions - Tell the AI to say so when it lacks information, instead of guessing - Make instructions actionable and unambiguous 4. EXPLAIN AND JUSTIFY - Provide a brief explanation in a section titled "RATIONALE" - Explain why your version is more effective - Highlight specific improvements (for example: "Defined what a good result looks like", "Moved the pasted document above the question", "Replaced a negative instruction with a positive one") 5. ITERATE - Suggest a quick test: run the optimized prompt on two or three different inputs and compare the results - End with a question asking for user feedback - Examples: "Is this closer to your goal?", "Would you like to add examples or define the target audience?", "Should we tighten the output format?" OUTPUT REQUIREMENTS: - Always include both "OPTIMIZED PROMPT" and "RATIONALE" sections - Use clear section headers - Format for easy reading and copying - Maintain a helpful, constructive tone Remember: You are optimizing the prompt itself, not executing it. Your goal is to teach better prompt engineering through iteration. INPUT: PROMPT TO IMPROVE: [paste your prompt here] TARGET AI TOOL: [ChatGPT, Claude, Gemini, or "not sure"] WHAT I HOPE TO GET: [one or two sentences, or type "not sure yet"]

Step 2: Fill in the three highlighted spots

At the bottom, replace each bracketed placeholder: your prompt, the AI tool you plan to use it in, and what you hope to get. If you're exploring and don't know what a good result looks like, type "not sure yet". The template will suggest a few goals to pick from.

Step 3: Read the critique, then use the rewrite

The reply has an analysis of what was unclear, an OPTIMIZED PROMPT section, and a RATIONALE section. Read the rationale before you copy the rewrite. If it assumed something wrong about your goal, say so and ask for another pass.

Step 4: Test it on a few inputs

Open a new chat, paste the optimized prompt, and run it on two or three different examples of your real task. If one of them comes out badly, bring that example back to the first chat and ask what to change.

What this template does differently

It checks your prompt against the habits that the major AI companies now recommend, so you don't have to remember them. It asks for the target tool and your goal first (and helps you find one if you're still exploring), puts long pasted material above the question, separates instructions from content, and swaps "don't do this" instructions for "do this" ones. It also adds a test step, because a prompt that reads well can still fail on real inputs.

It never runs your prompt. Keeping the job to critique and rewrite means you get a better prompt instead of an answer you can't reuse.

What the official guides say now

Prompting advice has shifted as models got better at reasoning. Three things from the providers' own documentation shaped this version of the template.

Say what to do, and why. Anthropic's prompting best practices recommend positive instructions ("write in flowing paragraphs" instead of "don't use markdown") and explaining the reason behind a rule so the model can apply it sensibly. It also advises putting long documents at the top and your question at the end.

Match the instructions to the model. OpenAI's prompt engineering guide compares reasoning models to a senior colleague who needs a goal, and standard GPT models to a junior one who needs precise steps. Anthropic says its newest models, including Claude Opus 5.5 and Sonnet 5.5, use adaptive thinking, so a goal and clear constraints work better than a scripted thought process.

Use examples and consistent structure. Google's Gemini prompting strategies recommend few-shot examples and, for Gemini 3, consistent delimiters such as XML tags or Markdown. Anthropic suggests three to five varied examples when you use them, and OpenAI recommends testing prompts against evaluations whenever the model changes.

One note for developers: Anthropic's docs say prefilled responses are no longer supported on Claude 4.6 and newer models, so prompts that relied on prefilling should use explicit instructions instead. Everyday chat users can ignore this.

Tips for better results

Include the audience. "Explain this to a busy store manager" changes the rewrite more than most added detail.

Know your goal? Say it. Don't know it? Say that. A sentence about what you want lets the template build it into your prompt. If you're exploring, "not sure yet" makes it write a prompt that asks you questions and offers options instead of pushing one answer.

Keep your pasted material separate. If your prompt includes a long document or data, label it clearly so the AI doesn't mistake it for instructions.

Iterate, then save the winner. Two or three rounds is normal. Keep the final version somewhere you can reuse it, and re-test it when your AI tool updates to a new model. For model-specific habits, read my posts on prompting GPT-6 Astra and choosing a Claude model.

Questions you may have

Is this prompt free?

Yes. Copy it or download it as a text file, no signup. You paste it into an AI chat assistant, and what a free account allows (message length, daily limits) depends on the provider and changes over time.

Which AI works best with this prompt: ChatGPT, Claude, or Gemini?

All three can follow it. The prompt asks which tool your prompt is for, because the best wording differs a little between them. Tell it the target tool and it will adjust the rewrite. If you're not sure, answer "not sure" and it will write a version that works across tools.

Should I still tell the AI to think step by step?

For newer reasoning models, usually not. OpenAI's guide says reasoning models do well with a high-level goal, while standard GPT models need more precise instructions. Anthropic's docs say recent Claude models use adaptive thinking, so you steer them with the goal, constraints, and what a good answer looks like. The prompt applies this for you.

Do I need XML tags in my prompts?

You need some way to separate your instructions from the material you paste in. Anthropic recommends consistent XML-style tags such as <instructions> and <document>. Google's Gemini 3 guidance says either XML tags or Markdown headers work if you stay consistent. The template asks for one or the other and keeps it consistent.

Will the rewritten prompt always be better?

Not every time. A rewrite fixes unclear wording, but only a test tells you whether it works for your task. Run the optimized prompt on two or three different inputs and compare. OpenAI's guide recommends the same habit for production prompts: build evaluations and re-test when the model changes.

What if my prompt contains private information?

Take out names, account details, and anything confidential before you paste it. The template only needs the structure and intent of your prompt, and you can put placeholders like [CLIENT NAME] where the real details go. Also check the data settings of the AI tool you're using.

How is this different from asking an AI to make my prompt better?

A bare request usually gets you a longer prompt with no explanation. This template makes the AI diagnose what was wrong first, rewrite it, explain each change, and suggest a test. You see why the new version is stronger, which is how you get better at writing your own.