How Unslop works

Before

It's worth noting that this pivotal update represents a significant step forward. We fixed the search bug. Your saved filters now work.

After

We fixed the search bug. Your saved filters now work.

The useful facts were already there. Removing the introduction lets the reader reach them sooner.

Most analysis uses ordinary code. A local grammar engine called Harper checks mechanics on the server. Rules look for particular wording. Word and sentence counts supply the readability and rhythm measurements. None of these checks calls a language model.

A flag means "look at this passage." A word like "tapestry" can be exactly right when you're writing about an actual tapestry. The checker does not know every exception, and an unflagged passage can still need work.

The style score runs from 0 to 100. Higher means more matched style patterns relative to the amount of prose. A score of 80 does not mean an 80% chance that AI wrote it.

Grammar, spelling, and punctuation scores run the other way: higher means fewer detected problems. The overall mechanics score averages those components. It does not grade the ideas, accuracy, or usefulness of the writing.

Rhythm deserves a separate look. Ten sentences can average the same length while sounding very different in sequence. Unslop measures both the spread of lengths and changes between neighboring sentences and paragraphs.

Uniform writing can be useful, especially in instructions or text for young readers. More variation is not automatically better. Unslop does not try to force a flat distribution or alternate short and long sentences.

The new rhythm and repetition findings do not change the style score. Their thresholds are provisional. They still need evaluation against approved human writing across different subjects and reading levels.

You can choose a reading grade, an approximate age from 6 to 18, or a maximum reading grade. Presets cover early readers through advanced adult text. The age option maps age to US grade by subtracting five, so age 10 targets grade 5.

An exact target accepts a range around the chosen grade. The default tolerance is one grade either way. A ceiling only sets an upper limit: a grade-8 ceiling accepts easier writing too.

The estimate uses the Flesch-Kincaid formula, based on sentence length and syllables per word. Unslop needs at least 30 words and two sentences to report a grade. It cannot establish whether a particular child understands the passage or whether its subject is suitable for them.

Writing complexity is reported separately as simple, moderate, or complex, using the same estimated grade. This describes the text, not the writer's ability. You can also tell the rewriting model who the audience is.

1. Unslop scores the original first. By default, AI editing runs only if the prose has at least 50 words and scores 20/100 or higher. You can change these cutoffs per request with rewrite_gate.threshold and rewrite_gate.minimum_words. Set rewrite_gate.enabled: false to allow edits regardless of score; the original is still scored and the bypass is reported. If the enabled gate fails, it returns the original, the score, and a skip reason without calling any model. This provisional cutoff selects text for editing; it does not identify its author.

2. Code divides the original prose into short, nonoverlapping spans, skipping protected content and prioritizing writing findings.

3. Calls alternate between OpenAI and Anthropic by default. Each model sees one editable span, nearby original context, and your voice guidance. It proposes one tiny replacement or leaves the span alone. Later calls never edit an earlier model's output.

4. Code checks each patch and the accumulated change budget, then assembles accepted patches at their original positions. It rejects adjacent patches that would combine into an oversized change.

5. The assembled text must meet all requested hard constraints, including reading targets and output length. Optional meaning review can add one model call to check the combined result. Failed or unverified constraints, or reported meaning changes, return your original.

By default, each replacement can remove and insert at most 8 words and 60 Unicode characters per side. The total change budget is 15%: added plus removed prose tokens, including punctuation, divided by the original token count. A word replacement spends both sides. Profile defaults and request constraints can change the total budget.

The default is up to four patch-generation calls, configurable up to twelve. Optional retries share that budget. Account allowance can reduce it further. The response reports coverage: spans left unreviewed stay unchanged. More calls mean more cost, not necessarily better writing.

Choose a single provider explicitly if you only want that company to process your text. The mixed setting needs both providers available. Each accepted patch identifies its provider and model.

Numbers, quotations, and common Markdown formatting are protected by default. You can protect specific terms or passages too. Micro-edits preserve paragraph breaks. Large reading-level or structural changes may be impossible within these limits.

There is no whole-document rewrite fallback. Invalid patches are discarded; an optional retry sees the original span and failure reasons. Provider errors fail the request without accepting an unchecked result. Small edits and provider diversity do not establish watermark removal.

For small corrections, the ordinary Proofread · no AI action applies only a limited set of mechanical fixes and your explicitly allowed replacements.

A saved profile holds your rules, editing instructions, and optional samples. You might prefer contractions, short introductions, or a particular technical term. A voice document supplies instructions; prose you've written supplies examples. They serve different purposes.

For descriptive voice comparisons, Unslop requires at least two samples with 20 sentences between them. It reports sample size and ranges for rhythm and contractions. If the declared genres differ, it withholds those ranges. They are not a voice-match score or rules that your next paragraph has to obey.

Some choices require judgment. An AI editor can help assess vague praise, repeated ideas, or a phrase that feels out of place. The deterministic checker can't fully evaluate them.

Analysis runs on the Unslop service hosted on Vercel. "Without AI" means no model call; it doesn't mean the text stays in your browser.

Rewriting and optional AI review send the relevant text and voice guidance to the selected model providers. Mixed editing uses both OpenAI and Anthropic. The service uses direct provider access when configured, or Vercel AI Gateway otherwise.

Prepaid requests may briefly cache response text for retries. See response retention and billing records. Saved profiles do store the instructions, rules, samples, and preferences you choose to save. Hosting and model providers have their own processing and retention terms.

External authorship detectors and watermark checks require explicit consent and a configured provider. No watermark verifier is enabled by default. Unslop doesn't remove watermarks, certify their absence, or promise to make writing undetectable.

Instruction-trap checks report suspicious patterns. They can't prove a document is safe or catch every possible instruction.

An API lets another program use these same tools. These groups cover the available operations. Every /v1 request needs authentication: an API key, or a same-origin signed-in session in the workbench.

The public OpenAPI reference gives the exact request fields. The API explorer lets people with a key try them. This explanation describes version 1.3.0, updated September 24, 2026.