Anthropic Just Told Everyone Claude Is Face Blind, On Purpose

Conventional wisdom says system prompts are trade secrets. Guard them like Coca-Cola guards its formula. Anthropic looked at that wisdom and shrugged.

Since July 2024, Anthropic has been publishing the literal system prompts that run behind Claude.ai and the mobile apps, right there in the release notes. Not a marketing summary. Not a vague "we care about safety" blog post. The actual text. And the page keeps getting updated — last touched August 14, 2026 — because these prompts change constantly and Anthropic just... tells you.

This is not normal. OpenAI doesn't do this. Google doesn't do this. Most vendors treat their system prompt like the recipe for the secret sauce. Anthropic treats it like a changelog.

What's actually in there is the fun part. When TechCrunch first dug into the July 2024 prompts, they found lines like "Claude cannot open URLs, links, or videos" and instructions for the vision system to behave as completely face blind. Not a technical limitation dressed up as one — an explicit policy choice, written in plain English, sitting in the prompt for anyone to read.

That's the real value here. If you've ever tried to reverse-engineer why a chatbot refuses something or formats an answer a certain way, you know the guessing-game misery. Anthropic just skips the guessing game.

The elephant in the room: none of this touches the API.

Anthropic is explicit about this, and it matters more than people give it credit for. The system prompts documented on this page apply to Claude.ai and the mobile apps only. If you're building on the API, you're not inheriting any of this — not the date injection, not the Markdown formatting rules, not the face-blindness disclaimer. Your production app and the chat window your non-technical coworker uses are running fundamentally different instruction sets on top of the same model.// That's a distinction I've seen developers miss constantly, and it's the kind of assumption that quietly breaks evals when someone tests behavior in the consumer app and expects it to match their API calls.<br>

There's also a structural shift buried in here that's easy to skim past: starting with Claude 4.6, each model ID is a fixed snapshot. One entry. No more dated revisions stacking up under the same model name. If you're doing regression testing or trying to pin down exactly what changed between versions, this is genuinely useful — Simon Willison already caught Anthropic quietly tweaking prompts between Opus 4.6 and Opus 4.7, and now that kind of diffing gets a lot cleaner.

Here's my honest take: this is a flex disguised as a transparency initiative, and I mean that as a compliment. Publishing your system prompt publicly is a confident move. It invites scrutiny. It lets competitors, researchers, and randos on Hacker News (this thread pulled 333 points and 144 comments, for what it's worth) pick apart your guardrails and call out anything hypocritical or half-baked.

But it also builds trust in a market where trust is the actual product. Enterprises buying AI tools want to know what's happening under the hood, especially around moderation and safety behavior. Anthropic handing over the instructions is worth more in procurement conversations than another benchmark chart.

Is there risk? Sure. Publishing the rules makes it marginally easier to probe around them. Some critics argue this exposes too much operational detail. Fair point — but I'd rather have a vendor that's occasionally exploitable-by-transparency than one that's opaque-by-default and asks me to just trust the brand.

Anthropic didn't need to do this. That's exactly why it matters.