What does the optimizer actually check for?

It scans the text with keyword patterns for each heuristic. Role detection looks for phrases like "You are" or "Act as"; context looks for words such as "background" or "our product"; instructions look for imperative verbs like "write" or "analyze"; format looks for "JSON", "markdown", or "table"; examples look for "e.g." or "for example"; and length simply tests that the prompt sits between 40 and 4000 characters.

Why does my well-written prompt score low?

The checker is pattern-based, so a prompt can be strong but miss the specific keywords a check scans for. If your role is implied rather than stated with "You are…", the role check fails even though the model might infer it. Read the failing checks and their hints, then decide whether to add the missing structure or accept the score — the checklist is a guide, not a verdict.

Can I use this to improve prompts for Claude or Gemini?

Yes. The six heuristics are model-agnostic: role, context, explicit instructions, output format, examples, and length improve results across GPT, Claude, Gemini, and open-source models alike. Nothing in the analysis targets a specific vendor, so the same score and suggestions apply regardless of which LLM you paste the final prompt into.

What happens to my prompts after analysis?

Nothing leaves the browser. The component analyzes the text in memory with local regular expressions and renders the score, checks, and suggestions in place — there is no API call, no upload, and no logging. You can paste confidential company drafts and client materials without exposure, and the tool keeps working fully offline once the page has loaded.

How do I fix a very long prompt warning?

When the prompt exceeds 4000 characters the length check fails and a suggestion asks you to split it or remove redundant detail. Look for repeated context, overlapping instructions, and examples that duplicate each other. Either trim to the essentials or split the task into a series of shorter prompts — the optimizer will show the length check pass once you are back inside the window.

Misc

Prompt Optimizer

Score and improve LLM prompts with actionable suggestions.

promptoptimizellmaiimprove

Loading tool…

About the Prompt Optimizer

The Prompt Optimizer scores any LLM prompt from 0 to 100 against six heuristics — role definition, context, specific instructions, output format, examples, and length — and explains exactly how to raise the score. Paste a draft into the Your prompt box and the tool grades each criterion with a tick or cross plus its point weight, then lists concrete suggestions such as adding a "You are…" role or naming an output format. A before/after panel contrasts a weak one-line prompt with a fully rewritten strong version you can copy with one click, which doubles as a worked example of the six heuristics in practice. Because the analysis runs entirely in your browser, no prompt ever leaves your device, which makes this AI prompt optimizer safe for confidential company drafts and client work. It is deliberately a checklist tool rather than a magic rewriter: it teaches the structure behind reliable outputs instead of generating text for you. For prompt engineers who want consistently better AI output without black-box tools, it turns vague intuition about good prompting into a measurable score you can iterate against.

Hand-written guide

Examples

Input

Write a blog intro about SEO.

Output

Score: 20/100 — Weak
Pass: Specific instructions (+20 pts)
Fail: Role definition, Context, Output format, Examples, Appropriate length
Suggestions (5):
1. Add a role definition — e.g. "You are a senior copywriter specialised in B2B SaaS."
2. Add context — describe the audience, product, or situation the model should consider.
3. Specify the output format — "Respond in markdown with H2 section headings."
4. Include 1–2 examples so the model understands the expected style and tone.
5. Your prompt is very short — add specifics, constraints, and expectations.

Note: The pre-loaded default. At 29 characters it falls below the 40-character floor, so five of six checks fail and the score lands at 20/100.

Input

You are a copywriter. Write a product description for a standing desk. Respond in markdown.

Output

Score: 70/100 — Needs work
Pass: Role definition (+20) · Specific instructions (+20) · Output format (+15) · Appropriate length (+15)
Fail: Context (15 pts) · Examples (15 pts)
Suggestions (2):
1. Add context — describe the audience, product, or situation the model should consider.
2. Include 1–2 examples so the model understands the expected style and tone.

Note: A mid-strength prompt: role, instructions, format, and length pass; context and examples are the gaps. The field hint reads 15 words.

Input

You are a technical recruiter. Context: we are hiring a senior React engineer for a fintech startup. Task: Write a job posting summary. Format: markdown. Example: "Build the core ledger UI."

Output

Score: 100/100 — Strong
Pass: all six checks — 20 + 15 + 20 + 15 + 15 + 15
Suggestion (1):
1. Strong prompt. Consider edge-case instructions ("If unsure, ask clarifying questions").

Note: All six checks pass, scoring 100/100 with the Strong label. The only remaining suggestion is the built-in edge-case nudge.

How to use

1
Paste a draft prompt into the Your prompt field (or start from the pre-loaded weak example "Write a blog intro about SEO.").
2
Read the Score card: the big number out of 100, the progress bar, and the Strong / Needs work / Weak label under it.
3
Review the six check cards — each shows a green tick or red cross plus its point weight and a hint for fixing it.
4
Apply the suggestions in the Suggestions card, starting with the highest-weight checks (role and instructions, 20 points each).
5
Compare your work against the Before / after panel, and click Use to copy the strong example — or copy the numbered suggestion list from the results box at the bottom.

Common use cases

Audit a shared prompt library — score each entry and fix the ones below 80 before rolling them out to a team.
Debug inconsistent model output: when replies drift off-format, run the prompt through the checks to find the missing format or examples.
Onboard new prompt engineers — the six check cards teach role, context, and format structure faster than a style guide.
Iterate before/after: score a prompt, apply the suggestions, and measure the score jump as a proxy for quality.
Rescue vague client briefs — paste the brief, add the missing role, context, and format elements the tool flags, and hand back a workable prompt.
Prepare prompts for API products where every call costs money — improve structure locally before spending tokens on test runs.

Best practices

Aim to pass all six checks before shipping a prompt: the checks map to the structure (role, context, task, format, examples) that consistently improves LLM output reliability.
Iterate in small passes — fix the failing checks with the highest point weights first (role and instructions at 20 points each), then re-read the score to see the jump.
Write examples that show the edge of your expectations, not the easy middle: one "do this" and one "avoid this" pair disambiguates style better than a single sample.
Keep prompts inside the 40–4000 character window the tool enforces: below 40 characters you are underspecifying, above 4000 the model dilutes attention.
Use Load weak example and Load strong example as training material when onboarding teammates — the score delta (20 vs 100) makes the six heuristics concrete.
Re-check length after every rewrite: shortening a prompt below 40 characters or pasting a huge spec can silently flip the "Appropriate length" check and drag the score down.

Tips

The score updates as you type, so fix one check at a time and watch the number climb — it is a satisfying feedback loop.
The weak example scores exactly 20/100 (only "Specific instructions" passes); use it to demo the tool to skeptics.
Copy the strong example with the Use button, then edit its specifics — it is a skeleton, not a finished prompt.
If you disagree with a failing check, read the hint: the checker looks for specific keywords, and sometimes your structure is fine without them.

Frequently asked questions

Six checks contribute weighted points: role definition and specific instructions are worth 20 each, while context, output format, examples, and appropriate length are worth 15 each. A pass on a check adds its weight; the total runs 0–100. Scores of 80 or more are labelled Strong, 50–79 Needs work, and under 50 Weak. Each check card shows a tick or cross with its point value.

Explore more misc tools

Browse the full collection of misc tools on the hub, or jump back to all categories.

All Misc tools All categories