Anthropic openly publishes the exact system prompts running under claude.ai and the Claude apps — and almost nobody reads them. The newest ones admit things most users have no idea about: silent safeguard routing to a different model, a restricted tier above the one you can buy, and named classifier warnings injected mid-conversation. This file is the 5 biggest admissions, every quote verbatim, with where each lives on the official page.

Get the free Admissions File

This is not a leak. Anthropic publishes the exact system prompts for every Claude model on its official docs (platform.claude.com → release notes → system prompts) — and the Claude 5-generation prompts contain admissions most subscribers have never heard: (1) ROUTING — by Anthropic's own published statement, queries on some topics (cybersecurity is their named example) are answered by Claude Opus 5 even when you selected Claude Fable 5; the safeguards 'sometimes catch harmless requests' and trigger 'on average, in less than 5% of sessions.' (2) THE MYTHOS TIER — Claude Mythos 5 and Claude Fable 5 'share the same underlying model,' but Mythos is 'not currently available to the public' — it runs inside Anthropic's Project Glasswing for a small number of trusted organizations. (3) INJECTED HISTORY — the June 2026 export-control suspension is written into the prompt as a dated notice, because 'Claude knows about them only from this notice.' (4) NAMED CLASSIFIER REMINDERS — image_reminder, cyber_warning, system_warning, ethics_reminder, ip_reminder, long_conversation_reminder — messages you never see, injected when a classifier fires. (5) THE PERSONALITY SPEC — the decline bar ('concrete, specific risk of serious harm'), the debate rule, and the famous 2025-era flattery ban. The file ends with 3 copy-ready prompt patterns drawn from how Anthropic writes for its own model, plus an honesty box on what none of this proves. Every quote verified verbatim against the live official page.

Subscribe to Hyperautomation AI ReportGet the PDF freeKeyword: ADMIT

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.