Write better prompts — seven rules, and a free scorer that checks yours against them
Prompt writing tips
Promptivo scores a prompt on seven dimensions, and each one comes down to a rule you can apply by hand. Here is one rule per dimension, shown as a real rewrite: the prompt before and after, with the score the engine gave that dimension each time.
Name the task and what it is about. A prompt built on "this" and "it" leaves the model guessing what you mean. Ask for one thing at a time, in plain sentences, and say what the answer should look like.
Look at this and tell me what you think of it and what to change.
Review my cover letter for a junior designer role. List the three weakest sentences as bullet points, and explain in one line why each one is weak.
"Short", "some" and "good" mean something different to every reader. Give the length in words, name the audience and the subject, and say what is in scope. Vague words such as "things", "stuff" and "maybe" count against a prompt.
Write a short summary of some things to know about saving for a pension.
Write a 150-word summary of how UK workplace pensions work for a 25-year-old in their first job, covering employer contributions and tax relief.
When the answer needs thinking, lay out the steps — first this, then that, finally the answer — in a sentence or two. Keep it short: the steps are the structure, and a paragraph asking the model to think hard only makes the prompt longer.
Which of these two laptops should I buy?
I need a laptop for photo editing and have £900 to spend. Compare the two models below step by step: first the screen, then the memory, then the price. Finally, recommend one in no more than 2 sentences. Do not repeat the full specs.
The model has not read your inbox or heard your earlier conversation. Put the role, the situation, the task and the form of the answer in the prompt itself — whatever you leave out, the model has to invent.
Reply to the customer email we discussed.
You are a support agent for our online bookshop. A customer, Sam, ordered three novels on 2 September, and one arrived with a torn cover. Write a reply that apologises, offers a replacement or a full refund, and asks for a photo. Return it as an email under 120 words, with the two options as bullet points.
"Nice" and "not too long" are wishes. A word limit, a fact that must be mentioned, a word that must not appear and a stated tone are rules you can hold the answer to. Lengths, counts, required and banned words, formats and tone all count.
Write a nice, not-too-long description of our water bottle.
Write a product description of our steel water bottle in under 80 words. Mention that it keeps drinks cold for 24 hours. Do not use the word "eco". Write in a friendly tone.
Say whether you want a table, a list or JSON, what goes in which order, and how long each part should be. Leave it out and the model picks the shape for you.
Give me some ideas for a team offsite.
Suggest 5 ideas for a one-day team offsite for 12 people, ranked by cost from lowest to highest. Return them as a table with a heading row: idea, cost per person, reason. Keep each reason under 15 words.
Name the source, the places and the dates, and say what a correct answer has to contain. A bare request for an opinion, with nothing to check it against, is marked down.
What do you think about electric cars?
Based on the IEA's Global EV Outlook 2024, summarise how electric car sales changed in China, Europe and the United States between 2022 and 2023. The output should give one figure per region and name the table it comes from.
All seven at once
It is the same request. The rewrite only says what the first version left the model to guess: who it is writing as, what to write, the rules to keep and the shape to return. Every dimension rises, and the overall score — the mean of the seven — goes from 2.13 to 4.30.
Score your own promptWrite something about our new app for social media. Make it good.
You are the marketing lead for Penny Jar, a budgeting app for UK students. Version 2.0 launches on 1 October with shared bills and a monthly spending report. Write 3 LinkedIn posts announcing the launch. First, pick one feature per post, then explain what it does for a student. Rules: - Each post must be under 80 words. - Mention "Penny Jar 2.0" once in each post. - Do not use the words "revolutionary" or "game-changing". Return the posts as a numbered list, separated by a blank line.
Every score in this section is Promptivo's own output for exactly the text shown, measured in September 2026. Scoring is deterministic, so the same text keeps the same scores until the engine itself is recalibrated.
Each dimension is scored from 1 to 5 on its own — the seven rules above, applied to your prompt — and the weakest come back with advice, lowest first.
Same prompt, same score, every time.
No round-trip to a model provider — the engine scores a prompt in a fraction of a second.
0 model callsScored on our own servers and never sent to anyone else. Prompts are kept as research data, so leave secrets out.
Four dimensions from the MePO framework, three inspired by Google Research's IFEval.
Why Promptivo
Free, permanently, and run as a Wizhut Research experiment: repeatable prompt evaluation grounded in linguistics and academic research — not another LLM grading another LLM.
Same prompt in, same score out — every time. No model drift and no run-to-run variance, so when you edit a prompt and score it again, any change in the score comes from your edit. Reproducible by construction.
No LLM in the loop at scoring time. Deterministic linguistic analysis feeds frozen, judge-calibrated scoring heads, so there is no model call to wait on — the engine needs a fraction of a second, even at the 50,000-character limit.
No AI model reads your prompt, and it is never sent to a third party — scoring runs on our own servers. Promptivo is a research experiment, so the prompts scored here are kept with their scores as research data. Leave secrets and personal data out.
Four of the seven dimensions come from the MePO framework (Zhu et al., arXiv:2505.09930, EACL 2026). The other three are inspired by Google Research's IFEval, which tests instructions a program can check.
We score our own work — and publish it: IFEval's 541 expert prompts average 3.09/5. The scoring was calibrated against one LLM judge, then checked on prompts it never trained on and against a second judge that contributed no training labels. Numbers, not vibes.
English gets the full linguistic analysis; Latin-script languages get structural signals; other scripts get honest structure-only scoring. The tiers are documented, not oversold.
FAQ · Writing better prompts
Not necessarily, and Promptivo does not promise it. The score measures how a prompt is written — how clear, specific and complete it is, and whether its rules can be checked — not what a model does with it. A prompt that scores well is easier to read, check and reuse; whether a particular model then follows it is a separate question, and no score here answers it.
No — the rewriting is yours. In the full report, every dimension that scores under 4 comes with a suggestion, and the three weakest are listed first, so you know which rule to reach for. The suggestions are written in advance for each dimension and score range, not generated for your prompt: they tell you what kind of change to make, and the words stay yours.
It is free, permanently: there are no plans and nothing to pay. Without an account you can score 5 prompts an hour, of up to 4,000 characters (fewer in non-Latin scripts), and see the verdict, all seven scores and the two weakest areas. A free account, once you verify your email, gets the full report, with written feedback on every dimension, plus prompts of up to 50,000 characters and a history of your last ten. There is no quota and no daily cap. Promptivo is a web app: there is no API, plug-in or browser extension.
They are kept. Promptivo is a Wizhut Research experiment, and the prompts scored here are its research data: each one is stored with the score it received, anonymous ones included, so the engine can be studied and recalibrated against prompts people actually write. No AI model reads them, and they are never sent to a third party. Leave secrets and personal data out. Your history is separate: it lives in your browser and is cleared when you log out.
Partly. English gets the full analysis. German, French, Spanish and other Latin-script languages get part of it, and other scripts — Greek, Russian, Chinese, Arabic and more — are scored on structure alone. Feedback is always in English, and the same prompt scores lower outside English, so compare scores within one language.
Not at scoring time. Each of the seven dimensions is scored from 1 to 5 by deterministic analysis, and the overall score is their mean. The standard it measures against was learned once from an LLM judge and then frozen, which is why no model sees your prompt and the same prompt always gets the same score. Four dimensions come from the MePO research framework and three are inspired by IFEval; the research write-up has the method and the measurements.
Free, permanently — no plans, no credit card. Paste a prompt, see the breakdown, take the advice. The seven rules above are what it scores, so its lowest score tells you which rule to reach for first.