Prompt Optimizer

The Prompt Optimizer applies deterministic, rule-based transformations to improve the clarity and structure of your prompt. It normalizes line endings and whitespace, collapses excessive blank lines, normalizes heading and list-marker style, detects recognized section headings, identifies template placeholders, and generates heuristic review suggestions for vague instructions, possible conflicts, missing output-format guidance, and repeated wording. Every change is visible before you copy the result.

No. The optimizer is entirely rule-based and runs in your browser using vanilla JavaScript. No prompt text is sent to a language model, a formatting API, or any server. You can verify this by opening your browser DevTools Network tab while using the tool — you will see zero outbound requests related to your prompt content.

A transformation is a mechanical change applied automatically when you enable the corresponding option, such as removing trailing spaces or normalizing a list marker from "* " to "- ". A suggestion is a heuristic review signal displayed in the Suggestions panel that identifies a possible issue without changing your prompt — for example, a vague verb phrase or a missing output-format section. You review suggestions yourself and decide whether to act on them.

The safe transformations are: normalizing CRLF and CR line endings to LF, removing trailing spaces from each line, trimming outer blank space, collapsing repeated blank-line runs to a configurable maximum (1, 2, or 3), normalizing spacing after ATX heading markers (##), normalizing list-marker spacing for unordered (- * •) and ordered (1) 1.) lists, and removing exact duplicate lines when that option is explicitly enabled. No substantive text is deleted or rewritten.

The heuristic suggestions are: Possible ambiguity (a vague verb phrase such as "make it better", "handle it", or "figure it out"), Possible conflict (two instructions that may contradict each other, such as "be concise" and "be very detailed"), Missing output format (the prompt describes a task but contains no explicit output-format signal), Possible repetition (a non-trivial line appearing more than once), and Detected section (a missing recommended section such as Role or Output Format). These are heuristic signals, not guaranteed findings.

No. The Prompt Optimizer uses local rules and pattern matching. It can improve presentation and flag possible clarity issues, but it cannot prove that a prompt will produce better results for any particular model. Model behavior depends on many factors beyond prompt structure, including model version, system context, temperature, sampling parameters, and retrieval context. Use the suggestions as a starting checklist, not as a correctness guarantee.

Fenced code blocks (``` and ~~~) and their internal indentation are protected before any transformation and restored afterward. Template placeholders such as {{name}}, ${topic}, [OUTPUT_FORMAT], and <variable> are never renamed, filled, deleted, or reinterpreted. URLs, Markdown links, quoted text, JSON, YAML, and XML embedded in the prompt are treated as text and are not subjected to prose-rewrite rules. All substantive content is preserved unless the user explicitly enables the duplicate-line removal option.

The tool detects {{double-brace}}, ${dollar-brace}, [UPPERCASE_BRACKET] patterns for known placeholder names like [OUTPUT_FORMAT] and [YOUR_VALUE], <angle-bracket> patterns for short variable names, YOUR_VALUE_HERE, and ENTER_X_HERE style placeholders. Detected placeholders are listed in the Placeholders panel and are always preserved unchanged in the output.

Section detection scans for lines that match recognized heading keywords case-insensitively. The recognized labels are Role, Persona, Goal, Objective, Context, Background, Instructions, Tasks, Constraints, Requirements, Inputs, Variables, Output, Output Format, Examples, Few-shot Examples, Evaluation Criteria, and Notes. Detected sections are shown in the Sections panel as a read-only outline. The optimizer never silently reorders content — section reordering is not available in the current release.

Yes. Given the same input text and the same selected options, the optimizer always produces the same output and the same suggestions. The transformation engine uses only regular expressions and string operations with no randomness. This means you can run the optimizer multiple times on the same input and always get the same result, which makes it safe to use in a review or quality-checking workflow.

No. Prompt text is never written to localStorage, sessionStorage, IndexedDB, cookies, or URL query parameters. The in-memory session history stores up to 8 recent optimizations that are visible only to you and disappear automatically when you close or reload the page. No prompt text appears in page titles, URLs, metadata, analytics, error messages, or logs.

Yes. Click Open file and select a local .txt, .md, .markdown, or .prompt file. The file is read by the browser FileReader API entirely in memory. No file data is uploaded to any server. The file text appears in the input panel and analysis runs automatically.

The Change summary tab shows a line-level diff between the original input and the optimized output. Lines that are unchanged appear in the standard text color. Added lines are shown in green with a + prefix. Removed lines are shown in red with a − prefix and strikethrough. Changed lines show both the original and the replacement. The first 80 diff lines are shown with a count of remaining lines if the input is very long.

When Remove exact duplicate lines is enabled, the engine scans all non-blank lines and keeps only the first occurrence of each unique trimmed line. Blank lines are passed through unchanged so paragraph structure is preserved. The number of removed lines is shown in the Statistics panel and in the downloaded report.

Swap to input copies the current optimized output into the input panel so you can run a second optimization pass on the result. This is useful when you want to apply transformations in stages or when you have used the output as a starting point for manual edits.

Download .txt saves the optimized prompt text as a plain-text file named optimized-prompt.txt. The file uses UTF-8 encoding and LF line endings. The download uses a browser Blob URL and never contacts a server.

Copy report copies a structured text report to the clipboard containing the applied options, input statistics, output statistics, transformation summary counts, all review suggestions, and the full optimized prompt. This report is useful for sharing, reviewing in a text editor, or pasting into a team document.

Yes. Once the page has loaded, all optimization, suggestion generation, copy, and download operations run entirely in your browser with no network calls. You can use it without internet access after the initial page load.

When enabled, this transformation ensures there is exactly one space between the # characters and the heading text (for example, "##Heading" becomes "## Heading"). It also removes trailing # characters from ATX-style headings (for example, "## Heading ##" becomes "## Heading"). The heading text itself is never modified.

This transformation standardizes the spacing after common unordered list markers. Multiple spaces after - or * are reduced to a single space. The bullet character • is converted to - for consistency. Ordered list items using the parenthesis notation "1)" are converted to the period notation "1.". This improves visual consistency without changing the list items themselves.

The Prompt Optimizer cannot evaluate whether a prompt will produce a correct, useful, or better result from a language model. It cannot detect semantic contradictions that require natural-language understanding, assess prompt quality for specific model architectures, measure factual accuracy, verify logical consistency, or predict model behavior. Model-backed rewriting is not available in the current browser-only release. All suggestions are heuristic signals that may produce false positives or miss genuine issues.

This suggestion fires when the prompt contains an instruction verb (such as "write", "generate", "summarize", "explain", "analyze", or "list") but does not contain an explicit output-format signal (such as "Output Format:", "respond in", "return only", or "format:"). Adding an explicit output-format section often reduces variation in model responses, but the suggestion is a heuristic and may fire on prompts that have implicit format expectations.

Not in the current release. A model-backed rewriting mode would require an approved backend or user-supplied model connection, explicit privacy, cost, provider, retention, rate-limit, and failure-state disclosures, and evaluation fixtures to test whether rewriting preserves intent. This feature is deferred and will not be added until those requirements are met. The current tool does not include a fake "AI-powered" mode that performs only local regex transformations.