How WriteMyAIPrompt works
WriteMyAIPrompt compiles your description into a prompt in six deterministic stages, then scores the brief you gave it against a six-dimension rubric. Nothing is hidden: the weights, the stages, and the exact wording of everything the engine adds are all published on this page, generated from the same source the compiler runs.
The six stages
- 1
Normalise
Request framing is stripped — "can you please help me write" becomes "write" — the core action verb is recovered, and the brief is measured for how much it actually specifies.
- 2
Classify
The task is scored against a keyword, verb, and phrase table and assigned to one of twelve archetypes. The table is a plain data file, not a model, so a wrong classification can be corrected in one line.
- 3
Order the sections
Each archetype has its own section order. Context is placed before instructions so the model reads the material before being told what to do with it, and the task is restated last where models weight it most heavily.
- 4
Render for the target model
The same content is written in the syntax each model follows best: XML tags for Claude, markdown headings for ChatGPT and Grok, a front-loaded task for Gemini, a comma-delimited descriptor stack for image models, and labelled plain text for everything else.
- 5
Quality passes
Uncheckable qualifiers are flagged and given measurable replacements, the requirements your task implies are made explicit, duplicated instructions are removed, and a thin brief gains an instruction to ask questions rather than invent details.
- 6
Score
The brief is scored 0–100 against the six published dimensions below, and every suggested fix is measured by actually rescoring rather than estimated.
The scoring rubric, in full
Six dimensions, weighted to 100 points. Every dimension is non-decreasing: filling in a field can never lower your score. The vague-word penalty applies only to the task itself, so adding a constraint that happens to contain the word “good” gets flagged but not punished.
| Dimension | Points | What earns them |
|---|---|---|
| Specificity | 25 | How precisely the task is stated: its length, whether it names concrete details such as numbers or proper nouns, and whether it leans on vague qualifiers. |
| Context | 15 | Whether the model is given the background and the intended reader it needs. |
| Output format | 20 | Whether the shape of the answer is defined: format, output type, length, and tone. |
| Constraints | 15 | How many explicit rules and exclusions bound the output. |
| Examples | 15 | Whether the prompt shows the model what a correct answer looks like. |
| Success criteria | 10 | Whether the prompt states how the output will be judged. |
Bands: 0–39 weak, 40–64 fair, 65–84 strong, 85–100 excellent.
The twelve task archetypes
Every prompt is classified into one of these. The archetype decides the section order, the default role, and which unwritten requirements get made explicit.
- Writing
- Producing original prose: articles, essays, stories, letters, and long-form copy.
- Coding
- Writing, reviewing, debugging, or refactoring software.
- Analysis
- Evaluating evidence, comparing options, and reaching a defensible judgement.
- Research
- Gathering, sourcing, and synthesising information on a topic.
- Image
- Describing a visual result for an image generation model.
- Data extraction
- Turning unstructured text into structured, machine-readable output.
- Conversation / roleplay
- Sustained in-character dialogue, simulation, or interactive practice.
- Teaching / explanation
- Making a concept understandable to a specific level of reader.
- Marketing copy
- Persuasive copy written to produce a specific action.
- Planning / strategy
- Turning a goal into an ordered, actionable sequence of steps.
- Editing / rewriting
- Improving existing text without changing what it means.
- Agentic / tool use
- Instructions for a model that plans, calls tools, and acts over multiple steps.
Words the engine flags
These 34 words read as requirements but cannot be checked, so a model cannot tell whether it has satisfied them. When one appears in your task, the compiler flags it and adds a measurable clause — though defining it yourself will always beat the generic replacement.
engaging · compelling · interesting · high-quality · good · great · best · better · professional · modern · clean · creative · effective · comprehensive · detailed · some · a few · several · various · short · long · brief · concise · in-depth · appropriate · relevant · optimal · robust · user-friendly · seamless · cutting-edge · etc · and so on · asap
Common questions
- Does WriteMyAIPrompt send my prompt to an AI?
- Not for the main tool. The compiler is a deterministic program that runs entirely in your browser, so nothing is sent anywhere. The optional "Polish with AI" button is the one exception, and it is clearly labelled: it sends the compiled prompt to DeepSeek to sharpen the wording. Everything else works offline.
- Why does my score stay the same after the tool rewrites my prompt?
- The score rates the brief you supplied, not the prompt the tool produced. Restructuring your words adds no information you did not give, so the number does not move. Raising it requires adding something real — context, an example, a constraint — which is exactly what the suggested fixes tell you to do.
- Is the scoring rubric public?
- Yes, in full, on this page. The six dimensions and their exact weights are listed below, and the page is generated from the same source file the compiler uses, so the published rubric cannot drift from the one being applied.
- Why is it free?
- Because it costs almost nothing to run. The compiler is a pure function that executes in your browser, so there is no per-user server cost and no API bill no matter how many people use it. That is an architectural decision, not a promotional offer.
- What are the twelve task archetypes?
- Writing, Coding, Analysis, Research, Image, Data extraction, Conversation / roleplay, Teaching / explanation, Marketing copy, Planning / strategy, Editing / rewriting, Agentic / tool use. Each one carries its own section ordering, default role, and the requirements that kind of work implies but people rarely write down.
Try it on the prompt builder, paste an existing prompt into the improver, or browse the library by category.