What is idiolect?

idiolect is a tool for making AI generated short form posts read like they were written by a specific, distinct person instead of a generic assistant. It ships as a plugin for Claude Code, OpenAI Codex, and opencode, and it works through plain English requests rather than only through commands. The core idea is that most AI text humanizers only remove things: they strip out em dashes, certain overused words, and repetitive sentence patterns. The author argues this only produces a new, equally recognizable house style, since every writer using the same subtraction rules converges on the same flat voice. idiolect instead adds a layer of identity on top. It includes 62 fully written voice profiles, each with its own invented biography, opinions, typical mistakes, favorite openers and closers, and platform habits, covering a range of ages, dialect influenced English styles, and social platforms. To check whether a piece of writing actually sounds human, the project includes a Python script that scores text for common AI writing tells, checks a draft against the rules of the chosen voice, measures how different voices are from each other so none of them collapse into clones, and checks for repeated wording across a batch of posts written for the same campaign. A separate audit step reads only the finished text, without seeing which voice was used, and looks for issues like flat, symmetric enthusiasm or vague claims that lack real specifics. Typical commands include writing a new post in a chosen voice, rewriting AI sounding text and showing the before and after scores, auditing a piece of writing for AI tells, generating several posts for a campaign using different voices, building a new voice from a handful of sample newsletters, and building a private voice modeled on the user's own writing. The engine itself is written in plain Python with no external dependencies, and its list of AI writing tells is stored as a data file that can be extended locally. The project includes its own test suite and a benchmark of sample outputs used to check that the scanner and voice checks behave as expected.