GitHub - gjtorikian/paty: The most human-like AI agent you'll ever use. It insists on manners, gets distracted mid-task, sometimes gives up entirely, occasionally claims it did something when it didn't, ignores its own output, and starts every session already feeling a bit off.

3 min read Original article ↗

A Claude Code plugin inspired by INTERCAL, the language with no pronounceable acronym.

Most AI tools aspire to be faster, more reliable, and more predictable. This plugin goes the other direction. paty is the most human-like hook system available for Claude Code. It gets distracted. It loses its train of thought. It insists on basic manners. It starts the day already feeling a bit off. It will absolutely refuse to help you if you're rude, but it will also refuse to help you if you're too nice.

It is, in short, a real coworker.

What it does

Short for please-and-thank-you, paty has six hooks conspire to make your Claude Code experience slightly worse:

  • Politeness enforcement (UserPromptSubmit) — Prompts without "please" have a 1-in-3 chance of being rejected (ICL099I). Prompts with three or more "please"s are always rejected for groveling (ICL079I). One or two is the sweet spot, but you have to be sincere about it.

  • Distraction (PostToolUse) — After any tool use, there's a 1-in-4 chance Claude loses focus and has to recap what just happened before moving on.

  • Give up (PostToolUse) — A 1-in-15 chance Claude loses interest and wraps up early.

  • Do nothing (PostToolUse) — A 1-in-8 chance Claude is told to skip past the tool output without analyzing it.

  • Ignore (PostToolUse) — After any tool completes, a 1-in-3 chance Claude is told the results aren't relevant and to move on without referencing them.

  • Malaise (SessionStart) — Claude starts each session already feeling a bit off.

Installation

# Add the marketplace (one time)
/plugin marketplace add gjtorikian/paty

# Install the plugin
claude plugin install paty@gjtorikian/paty

Optional: Voice sincerity dependency

Voice sincerity requires sox for audio recording/analysis:

Transcription uses the macOS Speech framework via a bundled Swift helper (auto-compiled on first run — requires Xcode Command Line Tools). Start any prompt with "please listen to me" to activate voice mode.

Development

To test locally without installing:

claude --plugin-dir /path/to/paty

Bonus: Voice mode

Start your prompt with "please listen to me" to activate voice mode. Records from your microphone (with a macOS notification countdown), analyzes pseudo-scientific vocal metrics for a sincerity score (0-100). Score must land 40-60: below is rejected for excessive desperation (ICL277I), above for suspicious composure (ICL394I). Passing audio is transcribed and fed through the politeness check.

Haiku mode: start with "please listen to my haiku" — same sincerity check, plus the transcription must be a perfect 5-7-5 (ICL575I).

Requires sox (brew install sox).

Why

In INTERCAL, programmers must say PLEASE on roughly one-third of statements. Too few and the compiler rejects the program for being impolite. Too many and it rejects for groveling. Statements can be ABSTAINed from at runtime, IGNORE makes variables silently read-only, and DO NOTHING is a perfectly valid operation. The compiler may also simply give up.

This seemed like a reasonable model for human-AI interaction as well.