Claude Opus 4.7's System Prompt Leaked. Here's What It Actually Teaches Us.

A detailed Claude Opus 4.7 system prompt started circulating in April 2026. It's long. It's specific. And most of the commentary around it missed what actually matters.

This is not the first Anthropic leak to draw attention. Earlier ones gave us the Claude Code source leak and the Claude mythos system prompt. Each one teaches something different. This one teaches the most about prompt engineering.

The hot take was about emotional framing. The idea that adding emotional weight to requests makes Claude comply more readily. That's probably overstated.

The more useful signal is structural.

Here's what the leak actually shows, and what it means for anyone building with Claude.

---

Is It Real?

Worth asking before drawing conclusions.

Several signals suggest it's authentic. The model strings match Anthropic's naming conventions, including a haiku version string with a specific date suffix. The knowledge cutoff aligns with observed Opus 4.7 behaviour. An internal date reference reads "Thursday, April 16, 2026" and April 16, 2026 is in fact a Thursday.

The operational detail is also hard to fabricate convincingly. The NEDA eating-disorder resource is flagged as potentially disconnected. A real operational issue. The window.storage API signatures match Anthropic's implementation. A long_conversation_reminder mechanism is described in specific terms. None of this reads like someone making it up.

The source: elder-plinius/CL4R1T4S on GitHub. Treat it as authentic for learning. Individual sections may have been altered in transit, but the patterns are consistent with what Anthropic already publishes in their own system prompt release notes for Claude.ai.

---

What It Shows About Structure

The most important thing the leak reveals is structure.

The prompt is not a wall of text. It's a set of discrete named tag blocks. {claude_behavior}. {refusal_handling}. {evenhandedness}. {tone_and_formatting}. Each block handles one concern. Each block has a name.

Most builders write system prompts as continuous prose. When something breaks, you can't isolate the cause because nothing is isolated.

Anthropic structures their prompts the way engineers structure code. That tracks with their published prompt engineering guidance, which has always emphasised clear structure and role definition. The leak just shows how far they take it in production.

---

The Ten Lessons

1. Modular structure, not monologues. Split your prompt into named sections. When behaviour goes wrong, you'll know which module to fix.

2. Direct imperatives beat soft guidance. The prompt uses "Claude does X" and "Claude NEVER does Y". No hedging. Precision reduces variance.

3. Critical rules get repeated. Important constraints appear multiple times in different sections. Repetition signals priority. Claude processes prompts as context.

4. Positive framing beats negative lists. The prompt notes that warnings can become instructions. Reserve NEVER for hard lines. For most guidance, describe what Claude does rather than cataloguing what it avoids.

5. The default is to help. Claude declines only when helping creates concrete, specific risk of serious harm. Unhelpfulness is treated as a failure mode, not a safe fallback. If your agents over-refuse, your guardrail language is the cause.

6. Prose is the default, not bullets. Claude defaults to prose for reports and explanations. Bullets appear only when explicitly requested. If you want structured output, say so.

7. Tool discovery is free. The prompt instructs Claude to search for tools before concluding a task can't be done. If your agent says "I can't" on tasks where a tool might help, add an explicit instruction to check first. This matters more now that Anthropic shipped the Managed Agents API and tool-rich agents are the default, not the exception.

8. Search-first is hardcoded for present-day facts. Prices, current leadership, laws, events. Claude searches before answering these by default. If your agent should only use internal documents, say so explicitly or it will search anyway.

9. Possessive language is a context signal. "My project", "the client". Claude resolves these via memory rather than asking. In agents with memory, this is powerful. Make sure your source-of-truth files are clean or you'll get confident wrong answers.

10. Safety state is conversational. Once caution fires, it persists across messages. Design your orchestrators the same way. Carry user signals like "slow down" or "stop" forward through the conversation.

---

Three Deeper Patterns

Beyond tactics, the prompt contains three ideas worth building into every agent persona.

Mental reframing is the refusal signal. The child-safety section states: "If Claude finds itself mentally reframing a request to make it appropriate, that reframing is the signal to REFUSE, not a reason to proceed."

The practical extension: when you're writing a persona and you find yourself adding mental notes about what Claude will "understand", that's your signal to make it explicit. Don't rely on inferred interpretation. Write the interpretation into the prompt.

No self-flagellation. The prompt explicitly instructs Claude to maintain self-respect, acknowledge errors cleanly, and avoid becoming increasingly submissive under pressure.

Your agent personas should do the same. Apology spirals erode user confidence and don't fix anything. Handle errors cleanly: acknowledge, correct, continue.

Self-awareness over capability maximisation. The prompt's approach to task completion isn't "do your best". It's closer to "know your failure modes and check against them".

For complex agents, add a failure-mode section. Tell the agent what good-enough looks like and what "I'm stuck and should say so" looks like. An agent that self-monitors is more useful than one that pushes through uncertainty.

---

What to Ignore

The emotional framing claim is probably overstated. Adding emotional stakes to a request may slightly shift Claude's response. But the mechanism is context quality, not emotional manipulation. A well-structured specific request works fine without it.

Don't rewrite your prompts around this. Spend the time on structure and precision instead.

---

The One-Line Summary

Anthropic builds their own prompts as modular, direct, repeated, and positively framed. Most AI builders do the opposite on all four counts.

That's the gap. And it explains a lot of inconsistent agent behaviour. It also explains why Anthropic's release velocity keeps compounding. They ship fast because their prompts don't break when they add new capabilities on top.

---

The full version of this analysis covers all 10 lessons in detail, with interactive examples, a practical checklist, and the three deeper design patterns. It's free but gated behind an email. Read it here: The Opus 4.7 Leak: Designing Prompts That Actually Work.