Humanize

Humanize ships as two LLM skill files — `humanize` and `ai-check` — each containing static, research-grounded instructions that any compatible agent executes on demand. The core workflow is one install command that places the skills in the right directory for your agent, then a natural-language invocation: you point the skill at a piece of text, and the agent applies the rule set. No external API call leaves your environment. No account is required. The skills are MIT-licensed and readable in plain text, so every transformation rule is inspectable before you deploy.

The differentiating feature is the sourcing. Rather than heuristic rewrites or prompt-engineering folklore, the `humanize` skill models nine levers — syntactic variation, burstiness, lexical entropy, and related dimensions — drawn from detection literature spanning peer-reviewed sources through April 2026. The `ai-check` skill inverts this: it scores incoming text and quotes the specific phrases triggering AI-signal flags, giving you evidence rather than a black-box pass/fail score. The vendor-published benchmark describes 25 inputs evaluated against two independent scorers, though reproduction depends on the agent and model you run it in.

The tool fits teams that want humanization logic embedded in an existing LLM agent workflow without standing up a separate service, and researchers who need a citable, auditable rule set rather than a proprietary rewriter. It breaks when the target detector updates its scoring model after April 2026, when you need bulk processing through a REST endpoint, or when your pipeline requires structured JSON output with confidence scores — none of which the static skill architecture provides.

Installation paths differ per agent — the README documents separate commands for Claude Code, Codex CLI, ChatGPT desktop, Gemini, Cursor, Aider, OpenCode, Continue, and Copilot — but the skill content is identical across all of them. The `install.sh` script handles path resolution; the skills themselves are plain instruction files, which means updating them is a file replacement, not a dependency upgrade.