Prompt Optimizer

A prompt optimizer that explains every change

This prompt optimizer reads your prompt the way a careful editor would and lists what to fix: vague requests (“help me with”, “make it better”), filler that costs tokens without adding meaning, rules that only say what to avoid, instructions repeated twice, instructions that contradict each other, and missing basics such as the audience or the output format.

Nothing changes until you say so. Each suggestion has Accept and Reject buttons, vague verbs come with replacements to choose from, and the page shows a before-and-after score on eight checks and the exact number of tokens saved, counted with OpenAI’s o200k_base tokenizer. An optional on-device AI can polish the result afterwards.

How to optimise a prompt

1. Paste your prompt into Your prompt.
2. Read the Suggestions: each shows the text it would change and how many tokens it saves.
3. Press Accept or Reject on each, choose a replacement for vague verbs, or use Accept all.
4. Compare the scores and token counts, and check the Changes view.
5. Fill in any [bracketed] parts, then copy the optimised prompt — or open Improve with AI for an on-device rewrite.

Prompt Optimizer features

Worked example

The example prompt asks for a product description, politely and at length:

Hi! Could you please help me with a product description for our new water bottle? It's for first-time hikers. It is very important that you make it really good. Don't use jargon. Don't make it boring. In order to stand out, mention the 24-hour cold promise and the 750 ml size. Don't use jargon. Keep it brief but detailed. Thanks in advance!

It passes 2 of the 8 checks and is 77 tokens long (o200k_base). Accepting every suggestion, with the first replacement for each vague phrase, gives:

Write a product description for our new water bottle. It's for first-time hikers. Make it clear and persuasive. Use plain, everyday words. Make it lively and engaging. To stand out, mention the 24-hour cold promise and the 750 ml size. Keep it brief but detailed.
Output format: [for example, three short paragraphs]

That is 69 tokens — 8 fewer, even with the new “Output format” line to fill in — and it passes 7 of 8. The check that still fails is deliberate: “brief but detailed” is a real contradiction, and only you can decide which one you meant.

Why these rules

Guidance checked 14 October 2026. The goal, format and conflict checks are shared with the AI prompt generator, so both pages judge those the same way.

• Specific beats vague. OpenAI’s and Anthropic’s prompting guides both ask for explicit instructions about the result you want; “help me with” leaves the model to guess.
• Say what to do. Anthropic’s guide recommends telling Claude what to do instead of what not to do — “Write in prose paragraphs” rather than “Don’t use markdown”.
• Fewer tokens, same meaning. Greetings, “Could you please…” and “Thanks in advance!” are paid for on every API call and every message; models need no politeness to comply.
• One instruction, once. Repeating a rule does not make it stronger, and two versions of a rule can drift apart when one is edited.
• No contradictions. A model asked for “brief but detailed” must pick one; better that you pick.

What it will not do

The optimizer does not judge whether your facts are right or your request is a good idea, and its rules work best on English prompts. It never removes names, numbers or {{variables}}. To rebuild a prompt into clear sections, use the prompt formatter; to count tokens for several models, use the LLM token counter.