Prompt optimization
Shorten and restructure a prompt without changing what it asks for. Every rewrite is checked against the original, and a rewrite that drops a requirement is rejected.
Most prompt optimizers hand your prompt to another model and return whatever comes back. That can read well and still lose a constraint. PromptFlowEngine takes the opposite approach: it applies a small set of documented transforms, then proves that the result still contains everything the original did.
There are three modes. Conservative normalizes whitespace and removes exact duplicate sentences. Balanced also removes filler and politeness phrases. Structured additionally arranges the prompt into Task, Context, Requirements, Output format, and Input sections and wraps template variables in tags.
Each mode produces a candidate. A candidate is kept only if every original sentence, every extracted requirement, and every protected span (code, quotes, URLs, variables) is still present, and it introduces no new warning. You choose a goal: fewest tokens, highest quality, or a balance. If nothing beats the original, the original wins.
Practical example
This run was produced by the engine when the page was built, in structured mode with the balanced goal.
Hello! I would like you to please rewrite the pull request description below so that it is clearer for reviewers. Please make sure to keep all of the technical details. Please make sure to keep all of the technical details. In order to help reviewers, add a 2-sentence summary at the top. Return markdown.
{{pr_description}}## Task
Rewrite the pull request description below so that it is clearer for reviewers.
## Context
To help reviewers, add a 2-sentence summary at the top.
## Requirements
- Make sure to keep all of the technical details.
## Output format
- Return markdown.
## Input
<pr_description>
{{pr_description}}
</pr_description>Computed by the engine when this page was built. Token change 5.9%; 3 issues fixed;0 left for you as suggestions. Structured mode may add headings, which is why the balanced goal allows a small token increase in exchange for a higher score.
Sentences, requirements, and protected spans are compared before and after. A failed check rejects the candidate and shows the reason.
Each edit is listed with its rationale, alongside a word-level diff of original against result.
Token counts before and after use the real tokenizer for OpenAI models, with the cost difference per call.
Gaps that need your knowledge, such as an unspecified audience, come back as suggestions instead of made-up text.
Conservative, balanced, and structured, from whitespace cleanup to full sectioning.
Optimize for tokens, for quality score, or for the best score within a token allowance.
Set a maximum and the optimizer reports whether a verified candidate fits inside it.
Code blocks, inline code, quoted text, URLs, XML-tagged blocks, and variables are restored byte for byte.
A separate operation that turns a one-line request into a sectioned engineering prompt, with every added line labeled.
Put two versions side by side and see the token, cost, and quality difference, plus which issues were fixed or introduced.
Run balanced mode with the tokens goal on a prompt that executes on every request, and take the saving with a guarantee about content.
Structured mode separates task, context, requirements, and output format, which makes the next edit much safer.
Share the diff and change list with a reviewer so the conversation is about the changes and not about spotting them.
Use a token budget to find out how far a prompt can shrink before something has to be cut by a human.
A model rewrites freely and may add, drop, or reinterpret instructions. This optimizer applies fixed transforms and then verifies that every original sentence and requirement is still present. It is less creative and much more predictable.
It removes known causes of poor answers, such as repetition, filler, and missing structure. It cannot measure answer quality, because it does not run the prompt. Test the result against your model for anything important.
Yes. The transforms are about wording and structure, which apply to any model. Token counts are exact for OpenAI models and approximate for other providers.
Code blocks, inline code, quotes, URLs, and template variables are masked before any transform runs and restored exactly afterwards.
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.