Prompt Optimizer

Paste a prompt and get a rule-based check. Nothing is uploaded. The tool does not call a model and it does not rewrite the prompt for you. It looks for the gaps that most often make a local model wander: no role, no format, a very short request, or a very long one.

How to use it

Paste the full prompt, including the source text if you want the length check to see it.
Click Analyze prompt. Read the notes under the box.
Edit the prompt in your own editor. Run the check again.
When the notes are quiet, try the prompt in Ollama, LM Studio, or Jan at a low temperature if you need a stable format.

What the checks look for

Length under about 40 characters. That is usually “summarize this” with no rules. The note asks you to add a task and constraints.
No role line. It looks for phrases like “you are,” “act as,” or “role.” A role is optional for a tiny extraction, and required when the standard of the answer matters.
No format. It looks for words such as bullet, list, JSON, format, exactly, or “return only.” If none appear, it asks you to name the shape of the answer.
Over about 300 words. Long prompts bury the rules. The note tells you to move rules to the top. It does not mean 300 words is a context limit.

If none of those fire, you get a single note: no major issues detected, and a reminder to test a lower temperature. That is not a guarantee the model will behave. It means the prompt is not missing the obvious pieces.

Examples

This will be flagged

make this email better

Too short, no role, no format. Add who the email is for, a word cap, and what must stay.

This should pass the obvious checks

You are writing a short reply. Task: Answer the question in 3 bullets. Rules: - One sentence per bullet - Return only the bullets Question: [paste]

Passing the check does not mean the three bullets will be correct. It means you told the model what shape to use. For the writing habits behind that, see How to Write Better Prompts for Local AI Models.

What this tool will not do

It will not score “quality” or style. Those are not reliable as keyword checks.
It will not detect a prompt that asks the model to invent facts. Read the prompt and add “use only the text below” yourself.
It will not see instructions inside an attached file. Paste the text.
It does not store the prompt. Refreshing the page clears it. There is no account.

A sensible order of tools

Start from a skeleton in the Prompt Template Generator or copy one from the Prompt Library.
Run it through this optimizer and fix the notes.
If you are pasting a long document, estimate size with the Token Counter before you hit the context limit.
Freeze temperature, then test. The method is in How to Get Consistent Results from Local Models.

Questions

Why didn’t it rewrite my prompt?

A rewrite would hide what changed. The notes tell you which piece is missing so you can add it in your own words. Templates you can edit are on the generator page.

It said there is no role, but I don’t want a role.

Ignore that note for simple jobs like “extract these two fields.” Add a role when the model is reviewing, drafting, or choosing a tone.

Can I trust a “no issues” result?

No. It only means the four checks above did not fire. Still read the output of the model.

Does this run a local LLM in the browser?

No. It is ordinary JavaScript on this page. Your prompt stays in the tab.