nagisanzenin/idiolect

Contribute to nagisanzenin/idiolect development by creating an account on GitHub.

Summary

idiolect is a plugin for Claude Code, OpenAI Codex, and opencode that ships 62 fully-specified voice profiles (fictional people with biographies, stance engines, style patterns, and systematic error classes) alongside a deterministic AI-tell scanner and a blind semantic auditor. Rather than just removing AI patterns from drafts, it installs a declared identity, then verifies the output against that fingerprint with a scan/conform/audit loop. The Python engine (scripts/idiolect.py) is stdlib-only and exposes commands like scan, conform, distance, overlap, pick, ledger, fingerprint, self, validate, selftest, and doctor.

What it is useful for

• Generating platform-specific posts in a chosen voice. A single brief ("promote a $15/mo booking tool") can be run through five different voice profiles and produce posts that the scanner rates 0–9/100 while an "AI control" version rates 84/100, with zero cross-post overlap.
• Auditing existing drafts. Asking "does this look AI-written?" runs the scanner (era-tagged tell lexicon + construction regexes + stylometrics) and a blind auditor that reads only the text to catch semantic tells regexes miss (no stakes, symmetric enthusiasm, hollow specificity).
• Humanizing. A draft can be rewritten against a voice and re-scanned, with a receipt like scan 5/100 (clean) · conform PASS · structure price/story/med.
• Campaign dedup. For N-post campaigns, the overlap and distance (roster anti-clone floor at 0.2) commands enforce distinctness across voices.
• Building a personal voice. The /idiolect:self command (Claude Code) keeps a local corpus of the user's prose-like prompts to construct a self voice, with drift reporting when the user's style shifts under AI influence. Per the README, nothing is uploaded; self off and self clear controls are provided.

How engineers can use it

Installation follows the documented plugin flow:

# Claude Code
/plugin marketplace add nagisanzenin/idiolect
/plugin install idiolect@idiolect

# Codex
codex plugin marketplace add nagisanzenin/idiolect && codex plugin add idiolect@idiolect
bash scripts/install-codex.sh

# opencode — add to opencode.json:
# { "skills": { "urls": ["github:nagisanzenin/idiolect"] } }
bash scripts/install-opencode.sh

After install, the natural-language interface is the documented entry point. Slash forms (/idiolect:write, humanize, audit, campaign, synthesize, self) and English requests both work; voices can be referenced as people — "the HVAC guy" resolves, slugs are for receipts. Custom voices live in $IDIOLECT_HOME/voices/, safe from plugin updates.

Stated limitations to note: lexical tells decay (the lexicon is era-tagged data, not scripture); the tool refuses astroturf-shaped requests (N unrelated "satisfied strangers") and includes no posting automation; and no "undetectable" claim is made — receipts report measurements, not guarantees.