Prompt analysis
Paste a prompt and see what a model would have to guess. The analyzer checks it against a published rule catalogue and lists each problem with the reason it matters and a way to fix it.
The prompt analyzer reads a prompt the way a careful reviewer would. It parses the text into sentences, sections, variables, and protected spans such as code and quotes, works out what kind of task it is, then runs every rule in the engine over it.
Each rule looks for one specific problem: no stated objective, vague words such as "some" or "stuff", sizes without numbers, instructions that contradict each other, a missing output format, repeated sentences, template variables that are not delimited, injection phrases, and secrets pasted into the text.
The result is a list of findings, a quality score out of 100 with a breakdown by dimension, an exact token count for OpenAI models, and an estimated input cost. Nothing is sent to a language model, so the same prompt always produces the same report.
Practical example
The prompt on the left was analyzed by the engine when this page was built. The findings on the right are its actual output, not a mock-up.
Write some code to handle user uploads. It should be fast and support a few file types, etc. Make it secure but keep it simple.
Computed by the engine when this page was built. Tokens counted with the gpt-4.1 tokenizer.
Every finding quotes the text that triggered it, explains why it affects the answer, and suggests a concrete change.
Rules are deterministic. Two people analyzing the same prompt see the same findings and the same score.
In the workspace the analysis runs in your browser tab. The prompt never leaves your machine.
There is no model call to wait or pay for, so the report updates as you type.
Clarity, specificity, completeness, structure, consistency, output specification, efficiency, and security.
Recognizes debugging, new code, refactoring, review, and other task types, so advice fits the task.
Lists the requirements it found in the prompt, which is a quick way to notice one you forgot.
Exact counts for OpenAI models with their own tokenizer; other providers are labeled approximate.
Each finding links to a page with a triggering example, the explanation, and a corrected prompt.
The workspace, extensions, MCP tools, CLI, and HTTP API all run the same engine.
Run the analyzer on a system prompt or template before it reaches production, and fix the warnings first.
When a model gives a different answer on every run, the cause is usually an unquantified size, a missing format, or a conflict. The analyzer names it.
Use the findings as review comments when onboarding people who are new to writing prompts.
Run every prompt file through the CLI and sort by score to find the ones that need attention.
No. The analyzer is a deterministic rule engine. In the web workspace it runs entirely in your browser, and it never calls a language model.
The findings apply to any large language model, because they are about the wording of the prompt. Token counts are exact for OpenAI models and approximate for Claude, Gemini, and others.
It tells you where the prompt is unclear, incomplete, or risky. It cannot run the prompt, so it does not measure answer quality. Use it to remove known problems before you test against a model.
Yes. The workspace needs no account, and the HTTP API needs no key.
Every product runs the same engine, so they combine without surprises.
Guides from the prompt engineering knowledge center that explain the ideas this product applies.
Free, deterministic, and private. No sign-up.