Prompt evaluation
A quality score you can check by hand. The score is 100 minus a fixed penalty per issue, every dimension lists the checks behind it, and two versions of a prompt can be compared line by line.
Evaluation answers one question: is this version of the prompt better than that one? A number from a model that will not explain itself does not answer it. PromptFlowEngine scores prompts with arithmetic that fits on one line.
Each finding subtracts a fixed penalty by severity. The overall score is 100 minus the sum, and it is capped when any finding is critical. Each of the eight dimensions is scored the same way from its own findings, and the report lists every check that ran in that dimension and whether it passed.
This is an evaluation of the prompt as written, not of model output. It tells you whether the instructions are clear, complete, consistent, and safe. Measuring what a model does with them takes a test set and model calls, which this engine does not make.
Practical example
The engine evaluated this prompt when the page was built. The table shows each dimension and the checks that failed in it.
Please write a short welcome email for new customers. It should be friendly and professional. Mention a few of our features and stuff. Keep it brief but include plenty of detail.
Computed by the engine when this page was built. 100 minus 29 points of penalties across 5 findings.
| Dimension | Score | Failed checks |
|---|---|---|
| Clarity | 90 | vague language |
| Specificity | 97 | unquantified size |
| Completeness | 100 | — |
| Structure | n/a | — |
| Consistency | 100 | — |
| Output specification | 87 | missing output format, missing length guidance |
| Efficiency | 97 | filler phrases |
| Security | 100 | — |
Every point lost maps to a named finding. There is no hidden model behind the number.
The score does not drift between runs or provider versions, so a change in score means a change in the prompt.
The same formula applies in the workspace, CLI, API, extensions, and MCP tools.
If you disagree with a rule, you can see it, ignore it in config, and say why.
Clarity, specificity, completeness, structure, consistency, output specification, efficiency, and security.
85 and above is strong, 70 to 84 good, 50 to 69 needs work, under 50 weak.
For each dimension, the list of checks that ran, which failed, and the text that caused it.
Quality, token, and cost deltas, plus findings resolved and findings introduced.
When optimizing, every candidate is scored and shown with the reason it was selected or rejected.
The engine exposes an evaluator interface; the built-in one is deterministic.
Compare them and read which findings each one resolves or introduces.
Record the score for each committed version and watch for regressions.
Pick a minimum score for production prompts and enforce it with the check gate.
The templates on this site are held to a strong score with no warnings, verified on every build.
The score starts at 100. Each finding subtracts a fixed penalty by severity: 30 for critical, 10 for warning, and 3 for info. If any finding is critical, the overall score is capped at 50. Each dimension is scored the same way from the findings in that dimension.
It means the prompt has none of the problems the rules look for. Answer quality also depends on the model, the task, and the input, which only output testing can measure.
No. The built-in evaluator is deterministic and makes no model calls.
Scores of 85 and above are graded strong. Short prompts for simple tasks often score lower on completeness and still work, so read the findings and not only the number.
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.