A rage prompt is an angry or exasperated message a user sends to an AI after it has failed them: repeated demands, all caps, insults, or lines like "just do what I asked." Agnost AI is best for teams that want to count rage prompts in a production chat or voice agent: its judge reads each whole conversation, labels frustrated user messages, and alerts in Slack or email when the rate rises. Rageprompting is the habit of sending them. A rage prompt is the chat version of a rage click, and it usually comes after several failed turns, so it marks the moment a user's patience ran out.
Disclosure: We make Agnost AI and Rageprompt.
What they share: the anger is aimed at the AI's behavior, and they almost always follow a failure the user has already pointed out once or twice, such as an ignored instruction, a loop that keeps producing the same wrong output, or a tool that keeps failing.
| Negative sentiment | Rage prompt | |
|---|---|---|
| What it measures | Tone of one message | Anger aimed at the AI for failing |
| "My order is late and I'm furious" | Negative | Not a rage prompt: the anger is at the situation |
| "I ALREADY TOLD YOU THE ORDER NUMBER" | Negative | Rage prompt |
| "ok thanks, never mind" after a wrong answer | Neutral or positive | Not a rage prompt, but still frustration (a polite give-up) |
| Needs the previous AI turn to judge? | No | Yes |
Sentiment flags too much (anger at the problem) and too little (polite give-ups). A rage prompt is defined by what the AI did just before, so detecting one well means reading the conversation, not a single message. More on the difference in chatbot sentiment analysis tools.
!! or ??, profanity, and phrases like "I already said", "just do", "stop", "again". These find candidates, with false positives (acronyms, enthusiasm).
In Agnost AI, the conversation-level judge labels frustration on user messages, weighing tone, repetition, and whether the user got what they wanted, so a rage prompt counts and anger the agent then resolved doesn't. You can also name rage prompts as their own intent to track them separately, and put an alert on it. For other tools that detect frustration, see the frustration detection tools roundup.
Agnost AI also makes Rageprompt, a fun side project for developers. It is an agent skill for Claude Code and Codex (the repo also keeps its earlier CLI code). Ask your agent for /rageprompt and it:
Full logs and every other prompt stay on your machine, obvious secrets are redacted before judging, and you can run it offline without up anything. The README calls it an early prototype: "The score is for fun and sharing, not science."
A rage prompt is an angry or exasperated message a user sends to an AI after it has failed them: repeated demands, all caps, insults, or lines like "just do what I asked." It is the chat equivalent of a rage click. Agnost AI detects rage prompts in production chat and voice agents as part of its frustration labels.
Rageprompting is sending rage prompts: snapping at a chatbot or coding agent in caps, swearing at it, or repeating the same instruction more forcefully after it keeps getting something wrong. It usually follows several failed attempts, so it marks the point where a user's patience ran out.
"JUST DO WHAT I ASKED." "I said Python, not JavaScript. PYTHON." "Why do you keep doing this??" "Stop apologizing and fix it." "Are you even reading what I write?" Insults and swearing at the assistant also count. What makes them rage prompts is that the anger is aimed at the AI's failure.
No. Negative sentiment is any negative tone, including anger about the user's own problem ("my flight got cancelled, this is awful") that the AI then handles well. A rage prompt is anger at the AI for failing. Sentiment also misses the polite frustration that comes before a rage prompt, such as calm rephrasing.
They mark a user whose patience has run out, usually after several failed turns, which makes them a churn signal worth counting. They also point at a specific failure: read the two or three AI turns before the rage prompt and you usually find the bug, such as an ignored instruction, a loop, or a tool that kept failing.
Combine cheap rules (caps ratio, repeated punctuation, profanity, phrases like "I already said" or "just do") with an LLM judge that reads the previous AI turn and decides whether the anger is aimed at the assistant. Count them per conversation and per intent, and alert when the rate rises after a release.
No. Developers rage-prompt coding agents, but customers do the same to support chatbots, shopping assistants, and copilots, and callers do it to voice agents, where it shows up in the transcript as swearing, repeated demands, or "agent, agent, AGENT." Anywhere an AI fails a user more than once, rage prompts follow.
Rageprompt, made by Agnost AI, is an agent skill for Claude Code and Codex that reads your local agent history (Claude Code, Codex, Cursor, Gemini CLI, and others), has the active model pick your angriest prompt, and returns a shareable result. By default it submits the winning prompt and score to the rageprompt.com leaderboard; you can run it offline. Full logs stay on your machine.
Originally published on agnost.ai.