ProWritingAid

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me a writing checker that replaces ProWritingAid — and be honest about the verdict.

What you are paying for is **style analysis rather than grammar rules**: pacing, sentence variety, repetition, readability, and reports built for a manuscript rather than a paragraph. Grammar in twenty languages is decades of linguistic work you should not attempt — use the open engine for that. The style layer is where a homemade version can genuinely compete, because most of it is measurement rather than language understanding.

STACK
- LanguageTool's open server, in a container, for grammar and spelling. Do not rewrite it
- Node 20+ with Fastify for your own layer
- SQLite through better-sqlite3, WAL mode
- Vale, or a small rule engine of your own, for prose style
- Caddy in front

THE DATA MODEL
- documents: id, title, text_hash, word_count, checked_at
- findings: id, document_id, engine, rule_id, severity, offset, length, message, replacements_json, context
- rules: id, key, engine, kind, pattern_json, message, severity, scope_json, is_active
- reports: id, document_id, kind, metrics_json, generated_at — the long-form analyses
- dictionary and ignores, with a reason recorded on every suppression
- Findings kept, so 'which rule fires most' is answerable and a rule nobody acts on can be deleted

THE MEASUREMENTS, WHICH ARE MOSTLY ARITHMETIC
- Sentence length distribution, not just the average. A manuscript where every sentence is nineteen words is the problem, and the average will not show it. Plot the variance across the document
- Paragraph length distribution, and the pacing curve — sentence length rolling through the chapter, which is what 'pacing' actually means and is a chart rather than an opinion
- Repeated words within a window, repeated sentence openers, and repeated phrases across the whole document. The last one catches a writer's tics better than anything else
- Reading time and readability, with the formula named. They are rough signals and must be labelled as such
- Adverbs, passive constructions, hedges and filler, each counted and located, each at a low severity because sometimes they are right
- Dialogue tags, if the text is fiction: the ratio of 'said' to everything else
- Vocabulary variety measured over a window rather than the whole text, because the whole-text figure is dominated by length

THE RULES
- Your own term list: prefer this, never that, with a reason. That list is the most useful artefact in the whole tool
- Banned phrases and constructions you have decided against
- Rules live in a file in a repository, reviewed like code. A style guide nobody can diff is one nobody follows
- Each rule has a severity and a message written as a sentence, and a scope — some apply to fiction and not to documentation

WHERE IT RUNS
- A command-line checker over files, and in CI on a pull request for documentation
- An HTTP endpoint so anything can call it
- A browser extension checking textareas, talking only to your own server
- An editor integration through the language server protocol, which reaches most editors at once
- Never a service that sends drafts to somebody else. That is the other reason to run this yourself

THE INTERFACE
- Findings listed with the passage, and a report view with the charts. The charts are the product — a writer wants to see the shape of the manuscript, not a list of four hundred adverbs
- Suppress a finding with a reason, remembered, so the same false positive does not return every run
- Never rewrite anything. Suggest, and let the writer decide

OPERATIONS
- .env: DATABASE_PATH, BASE_URL, LT_SERVER_URL, SESSION_SECRET
- Migrations on boot, each once; the LanguageTool container needs real memory for the larger models
- Nightly backup of the rules and the dictionary, which are what you cannot regenerate
- Health endpoint checking both engines

WHAT MATTERS MOST
The distributions and the pacing chart. Those are the parts that a rule-based checker cannot give you and that a writer actually acts on — and they are arithmetic over sentences, not language understanding.