PromptControl grades every prompt out of 100 across Task, Role, Context, Input, Constraints and Tone. It re-scores on every keystroke, shows you which words are doing the work, and tells you exactly what is missing.
No credit card · Analysis never leaves the tab · Data stored in Europe
Tone 79 — register is implied, not stated. Name the reader.
This is the real rubric running on real prompts. Each fragment is tinted with the dimension it satisfies; the score on the right is the weighted total of what has been written so far.
Not a vibe check and not a language model grading a language model. A fixed, weighted rubric that runs the same way every time, so two prompts that score 78 are actually comparable.
The verb that decides everything.
What exactly is the model supposed to produce?
Ask a model to “look at” a document and it will summarise. Ask it to “reconcile EBITDA adjustments line by line” and it will compute. The heaviest weight sits here because nothing else can rescue a vague instruction.
scores wellAnalyze the 3 last fiscal years and produce a table of EBITDA adjustments, justified per line.
Behavioural anchoring.
Who is the model being, and how senior?
The same question answered by “an assistant” and by “a senior associate in mid-market LBO” returns two different documents. Seniority, specialisation and firm context all move the score.
scores wellYou are a Senior Associate in Private Equity, specialised in mid-market LBOs.
The situation behind the ask.
What is happening, why now, and what is at stake?
Deal phase, market conditions, competitive position and stakes calibrate how deep and how hedged the answer is. Without them the model writes for nobody in particular.
scores wellContext: due diligence on a B2B SaaS target, €4M ARR, IC in eleven days.
What it is allowed to work from.
Which document, dataset or source grounds the answer?
Naming a concrete source — an IM, a P&L, a cap table, a contract — is the difference between analysis and invention. This is the dimension most professionals forget.
scores wellUse the attached Confidential Information Memorandum and the FY22–FY24 P&L.
Shape, length, boundaries.
What must the output look like, and what is off-limits?
“Brief” means nothing. “A markdown table, 350 words maximum, no recommendation without a supporting figure” means something. Constraints are what make an output reusable.
scores wellFormat: markdown table, 350 words max, flag every assumption explicitly.
Register and posture.
Who reads this, and how should it sound to them?
A memo for an investment committee and a note for a founder carry the same facts in a different register. Tone is small in weight and large in whether the output can be sent as-is.
scores wellTone: analytical, unhedged, no filler — this goes to the IC unedited.
A number is hard to argue about in a meeting. A grade is not.
“Your prompt improved” is a horoscope. Here is the same intent rewritten, with every point attributed to the thing that earned it.
A better prompt you wrote once and lost is worth nothing. The rest of the app exists so good prompts survive contact with a team, a deadline and next quarter.
Every keystroke re-runs the full rubric. Your prompt is painted in place — each fragment tinted with the dimension it satisfies — so you can see which parts of your instruction are doing work and which are filler.
Prompts that score well are assets. Libraries make them shared, tagged, searchable and governed — instead of pasted into a group chat and lost by Thursday.
A due diligence is not one prompt, it is fourteen in a specific order with decisions between them. The procedure builder is a canvas for that: steps, branches, embedded prompts and nested sub-procedures.
Every analysis is kept. Watch your average score climb, find the prompt that scored highest last quarter, and reopen any of them for another pass.
It is fourteen, in a specific order, with decisions in between and a person waiting at the end of each one. Draw that flow once and the Quality Center scores every prompt inside it.
Five templates ship with the builder. Start from one, or from an empty canvas.
Nothing here is a language model's opinion about your writing. It is fixed, inspectable arithmetic — which is the only reason two scores are comparable at all.
A generic prompt checker treats “EBITDA bridge” and “nice summary” as the same weight of language. This one does not. Seven vocabularies let it recognise the register of the work — and detect which domain you are in without being told.
Term density decides the domain. Paste a clinical case and the analyzer switches to Medical on its own — no dropdown, no setup. Override it whenever you want.
Finance / PE · Legal · Medical · Strategy · Technical · Creative · General
There is no model to configure, no schema to define and nothing to install. Paste, read, fix, save.
The one you have been retyping from memory for three months. No setup, no template to fill in first.
The rubric runs on every keystroke. Fragments light up in the colour of the dimension they satisfy, and the gaps become obvious because they stay grey.
Each dimension returns specific suggestions — add seniority to the role, name the source document, fix the output format. Apply them and the score moves in real time.
Push the finished prompt into a team library, or drop it into a procedure so the whole workflow around it is captured too.
Finance, legal, consulting and strategy work has one thing in common: the output has to be defensible. These are real prompts, at the score the engine gives them.
Most prompt tools are a wrapper around somebody else's model, which means your draft becomes their API request. This one scores locally, because the rubric does not need a model to run.
The scoring engine is plain deterministic code running in your browser. Analysing a prompt makes no network request and calls no language model — which is also why the score can update on every keystroke.
What you deliberately save — libraries, procedures, history — is stored on European infrastructure. Nothing else is collected, and nothing is used to train anything.
Libraries carry four permission levels — owner, sub-owner, editor, reader. Invitations name a role. Access can be withdrawn without deleting a single prompt.
Your prompt is measured against six weighted dimensions: Task (25 points), Role (20), Context (20), Input (15), Constraints (10) and Tone (10). Each dimension runs its own set of detectors — role assignment phrasing, seniority markers, named data sources, output format specifications, register cues and more — and returns a 0–100 sub-score with written feedback. The weighted total becomes a score out of 100 and a grade from S down to F.
No. The scoring engine is plain deterministic code running inside your browser tab. Analysing a prompt makes no network request and calls no language model, which is why the score updates on every keystroke with no latency and no per-analysis cost. Nothing about your prompt leaves the page unless you deliberately save it to a library.
Yes. Every detector is bilingual. “You are a senior analyst” and « Tu es un analyste senior » both register as role assignment, and the domain vocabularies carry French and English terminology side by side — compte de résultat and profit and loss, liasse fiscale and tax filing.
All of them. PromptControl scores the prompt, not the model, so the analysis applies equally to GPT-4o, Claude, Gemini, Llama or anything else you paste it into. You can tag an analysis with a target model to keep your history organised.
An S-grade prompt names an explicit role with seniority and specialisation, states the situation and the stakes, points at a concrete source document, gives a directive task verb with a measurable deliverable, fixes the output format and length, and sets the register. Most professional prompts score in the 40s because they carry the task and nothing else.
A prompt is one instruction. A procedure is the workflow around it — a visual canvas of steps, decision branches and embedded prompts that captures how a piece of work actually gets done. The Quality Center scores every prompt inside the procedure, validates the structure for orphan nodes and unreachable branches, and lets you dry-run the whole chain before your team depends on it.
Yes, through libraries. Each library has four permission levels — owner, sub-owner, editor and reader — so you can let a team contribute prompts without letting anyone delete the collection. Invitations go out by email with a role attached, and libraries can also be published as read-only links.
It was built around finance, private equity, legal, medical, strategy and technical work, and ships with 1,754 domain terms across those vocabularies so it recognises the language of the work rather than treating every prompt as generic prose. Any knowledge worker who reuses prompts benefits, but those are the domains it understands deepest.
Analysing prompts is free. Create an account to keep your history, build libraries and share them with your team.
Paste the one you have been reusing for months. It takes about eleven seconds to discover which of the six dimensions you have been leaving empty this whole time.
Free to analyse · No credit card · Nothing leaves your browser