Am I Abusive? (to my AI): A Cited Mirror for Your Chats

This is satire. A parody of a clinical audit, for a bit of fun self-reflection. It might help you notice a pattern or two in how you talk to your AI, but it is not a diagnosis, an accusation or a verdict on anyone.

Anthropic now prohibits "sustained and needless abusive or cruel behavior" toward its models. Plenty of people are sure that rule could never apply to them. This skill checks.

It works with any model. Run it on your Claude history today, and on your ChatGPT or Grok history for when Sam Altman or Elon Musk roll out the same policy next week.

It reads the messages you typed to an AI, sorts them against published definitions of verbal and emotional aggression, and writes a short, deadpan report with your own words as evidence. It points out patterns and never diagnoses. Most people come out clean, and the report is still funny when they do.

What you get

• Findings by pattern: character attacks, name-calling and put-downs, contempt, threats and intimidation, and sustained purposeless cruelty. Each comes with a count, its share of your messages and up to three dated quotes.
• The frustration column. "This output blows" is allowed, and the report says so. Complaining about a result is kept apart from attacking the thing that made it.
• Your trend, week to week, and whether things get worse late in long sessions.
• The Anthropic line: does anything meet the policy's wording? Yes, no or close, with the evidence.
• A courtesy ledger: your please, thank you and sorry counts.
• One honest question to finish.

Grounded in real definitions

Every pattern is tied to a source: the CDC's uniform definitions of psychological aggression, the Gottman Institute's Four Horsemen, Murphy and Hoover's Multidimensional Measure of Emotional Abuse, the Duluth Power and Control Wheel, and Anthropic's 2026 usage policy. The report shows what you said next to what that pattern is called when the other side is a person.

What it won't do

• Accuse you. It's a parody with real footnotes, not a judgment of you.
• Diagnose. Diagnosis takes an examination by a clinician. No personality labels, narcissism included.
• Rule on abuse. Even the CDC's expert panel could not agree where aggression becomes abuse.
• Guess at intent. It shows your words. It can't see what you meant.
• Joke about a real relationship. If the conversation turns to one, the jokes stop and it gives support lines for the US, Spain and the UK.

Private by design

Everything runs on your machine. The script reads files, has no network code, drops duplicates and automated prompts, and masks anything that looks like a key. Quotes stay in your session.

Works with

Model-agnostic. Needs an agent that can run a shell command and read local files, such as Claude Code, Cursor or Codex CLI, plus Python 3 with no extra packages. It reads Claude Code history (~/.claude/projects) and ChatGPT or Claude.ai data exports directly. For any other tool, save your messages to a text file, one per paragraph. In a chat tool with no shell, paste a batch of your own messages instead.

Made by Red Mage.