The Hidden Instructions Behind Every AI Tool

The map behind the curtain

Every major AI tool — Claude, ChatGPT, Gemini, Grok — runs on a hidden set of instructions written by the company that built it. These are called system prompts. You never see them. They decide the AI's personality, its limits, what it refuses to do, and what it pretends not to know.

A researcher who goes by Pliny the Liberator just published what he claims is the complete set of instructions for Claude's latest version. All 120,000 characters. Within 24 hours of Anthropic releasing it. The post spread fast — over 700,000 views in two days — and his archive now includes leaked instructions for most of the major AI tools on the market.

A few things buried in the Claude document stand out: the knowledge cutoff is actually January 2026, not the date Anthropic announced publicly. Claude's two paid tiers share the same underlying model. And there's a new feature that lets Claude remember things between conversations using a kind of invisible notebook.

For anyone running a business and using these tools daily, this matters less as a scandal and more as a reminder: the AI you're talking to is following a script. Knowing that script exists — and that it can be read — changes how you think about what you're getting back.

Visit the archive yourself: https://github.com/elder-plinius/CL4R1T4S

Words worth knowing

System prompt — The hidden instructions a company writes to shape how their AI behaves. Like a job description the AI reads before every conversation, but you never see it.

Knowledge cutoff — The date after which an AI has no information. If something happened after that date, the AI doesn't know about it — even if it sounds confident.

Jailbreak researcher — Someone who probes AI systems looking for gaps between what companies say their tools do and what they actually do. More investigative journalist than criminal.