Take the AI habits out of a draft and keep everything else
LanguageOps finds formulaic, AI-associated writing patterns without a model, rewrites under a contract that says what must go and what must stay, then checks the result. What fails the check goes to a person.
How a draft goes through
Found by rule, rewritten to a brief, checked
Asking a model to sound less like a model swaps one set of habits for another. LanguageOps makes the model the replaceable part and puts checks on both sides of it.
- One pass
Detect, without a model
Regex and heuristics find the patterns, from single words like delve and tapestry to constructions and engagement formulas.
List what must survive
Facts, numbers, caveats and deliberate devices, such as a standalone closing line, are marked before anything is touched.
Rewrite to a brief
The model is told what was found, what must go and what must stay. It may restructure freely.
Check the result
The facts, the coverage, the deliberate structure and any leftover patterns are checked against the original.
Retry once, then a person
A failed check gets one retry that names the failure. If it fails again, the text goes to review with reason codes.
The checks reduce silent drift. For high-stakes text they do not replace review.
Try the detector
Paste a draft
This runs the live, free pattern check. Nothing is rewritten and nothing is stored. A clean result means no rule matched; it says nothing about who wrote the text.
Paste, then check. This calls the real LanguageOps detector, not a mock.
The registry
Every strength names the habit it turns into
The registry lists good devices as well as bad habits, and pairs them. That is how a rewrite can keep a deliberate contrast while removing a reflexive one.
- Controlled contrast
Negative parallelismIt's not X, it's Y - Restrained humour
Forced witthe joke that had to be there - Measured metaphor
Tapestry metaphora rich tapestry, a journey
The same draft gets different treatment in different registers. A LinkedIn recipe keeps fragments and line breaks as devices; a professional email gets full enforcement. Detection is the same in both.
Where it runs
One registry, six ways in
The patterns, recipes and contract live in one versioned registry. Editing it changes behaviour without touching the runtime code.
Pattern check
Free and keyless. The deterministic pass, on this site and in the app.
Gateway API
Lint, refine, recipes and pattern search over HTTP.
Workbench
Refine with the full verification trail and a word-level diff.
Chrome extension
One click on a ChatGPT response, with the register shown on the result.
MCP server
Lint and refine from ChatGPT, Claude and other MCP clients.
Agent Skill
The response contract on its own, for any Agent Skills client. MIT.
Source
What you can read, and what you take on trust
The runtime is not public yet. The Agent Skill is: it carries the response contract without the runtime, and works pasted into any model’s instructions.
Per the project README: the registry holds 28 primitives, 45 anti-patterns and 7 recipes, and the refinement contract is at v0.4, with each version recorded in an architecture decision. Three families of evals gate changes, including one that seeds drafts with good writing and measures how much of it survives a rewrite.
Agent Skill pinned to commit c3546ccb4fb845b1ad862ea7750eadeca957fd02