What a Leaked Claude System Prompt Teaches You About Writing Better AI Instructions (Ignore the "Emotional Framing" Take)

In April 2026, a detailed Claude Opus 4.7 system prompt surfaced on GitHub. Within 48 hours, the internet had decided the lesson was "emotional framing is the secret".

It is not. The emotional framing is the distraction. The structural patterns are the real lesson, and they are what will actually improve your own AI prompts - whether those prompts are for SDR agents, a proposal-drafting skill, a CLAUDE.md file, or a system prompt on a Managed Agent.

Here are the ten patterns that matter, translated into how you should write prompts for your own team.

First: is the leak real?

Short version, yes. Long version, the signals are:

• Model strings match Anthropic naming conventions.
• The Haiku version string carries a date suffix consistent with Anthropic's pattern.
• The knowledge cutoff aligns with observed Opus 4.7 behaviour.
• Internal date references check out.
• Operational details like the NEDA safety-resource flag and window.storage API signatures are hard to fabricate.

Treat the leak as genuine with the standard caveat that individual sections could have been edited after the fact. The structural patterns are too consistent to have been invented.

The ten patterns that actually matter

1. Modular structure with named tag blocks

The prompt is not one wall of text. It is a series of clearly named tag blocks: one for tone, one for safety, one for output format, one for tool use and so on.

Apply this to your prompts: stop writing your system prompt as a single block of prose. Split it into sections. Name each section. Claude will treat each section as a distinct set of instructions instead of one mushy instruction soup.

2. Direct imperatives, not softened suggestions

The prompt says "do X". Not "you may want to consider X". Not "if possible, try to X".

Apply this to your prompts: strip hedging language out of your AI instructions. "Ask two qualifying questions before drafting a response" is a better instruction than "you might want to try asking a couple of qualifying questions". The model mirrors the directness of the prompt.

3. Critical rules are repeated

Important rules appear more than once in the prompt. Different sections, different phrasings, same rule.

Apply this to your prompts: for anything genuinely non-negotiable - compliance, tone of voice, escalation patterns - state it twice, in two different sections. Repetition is weighting. The model is more likely to honour rules that show up in multiple places.

4. Positive framing beats long don't-lists

The prompt describes what to do more than what not to do. When it does use a "don't", it is targeted and specific.

Apply this to your prompts: a prompt that lists 30 things not to do is a prompt full of ideas Claude is now thinking about. Say what good looks like. Only use "don't" for sharp, specific prohibitions.

5. Explicit output format

The prompt says exactly what format to use. Prose is the default. Bullets are only used when requested.

Apply this to your prompts: be explicit. "Respond in prose." Or "respond as a numbered list of three items." Implicit format guidance produces inconsistent output. Explicit format guidance produces consistent output.

6. Tool discovery and search-first behaviour

The prompt pushes Claude to search for context before answering, and to load tools on demand rather than assume they are all available at the top of the session.

Apply this to your prompts: if your agent has access to a CRM, a knowledge base and a calendar, tell it explicitly when to look things up vs when to answer from memory. Default to "look it up" for anything involving current state.

7. How safety state and authenticity are handled together

The prompt navigates the genuine tension between being helpful and being careful. It does not collapse to one or the other. Each safety rule has a scope and a reason.

Apply this to your prompts: guardrails need scope and reason, not just "be careful". "Do not promise pricing not listed on the sales sheet, because pricing requires legal review before it changes" is a useful rule. "Be careful about pricing" is not.

8. Default to helpful

The prompt pushes Claude toward being helpful. Over-refusal is treated as the common failure mode, and vague guardrail language is identified as the usual cause.

Apply this to your prompts: if your AI keeps refusing things it should handle, the cause is almost always a vague safety instruction. Tighten the scope. "Do not give medical advice" is a reasonable rule. "Be cautious about health topics" makes the model refuse half the inbox.

9. Possessive framing ("the user's...")

The prompt uses possessive framing frequently. "The user's request", "the user's intent", "the user's data". This is subtle but it consistently nudges the model toward treating the request as something it is serving, not something it is passively responding to.

Apply this to your prompts: frame the work as serving the user. "The prospect's qualification state" reads differently to "the qualification state". Small shift, measurable improvement in relevance.

10. Meta-lesson: the prompt describes the shape of the work, not the content

The final pattern is the one most people miss. The Opus 4.7 prompt does not specify answers. It specifies how answers should be shaped - format, tone, escalation, boundaries, defaults. The content comes from context at runtime.

Apply this to your prompts: write prompts that describe the shape of the work. The specific content - which prospect, which deal, which product - arrives at runtime. A prompt that tries to hardcode content ages badly. A prompt that describes the shape of good work ages well.

The take you should ignore

"Emotional framing is the secret."

The prompt does contain some emotional or identity-adjacent framing. Most of it is peripheral. Treating it as the main lesson is like reading a Michelin-starred recipe and concluding the secret is the chef's apron.

The structural patterns - named blocks, direct imperatives, repeated critical rules, positive framing, explicit output format, search-first behaviour, scoped safety, default to helpful, possessive framing, shape-not-content - are what actually make the prompt work. They are also what will make your prompts work.

Where to go next

For a deeper, section-by-section walkthrough with before-and-after examples for each pattern applied to SDR scripts, CRM AI rules, and proposal-draft agents, see our 10 Patterns guide. For the broader context on where prompts live in the Anthropic stack, see our Five Levels of AI in Revenue Ops guide and the Code vs Cowork decision framework.

TL;DR

• A detailed Opus 4.7 system prompt leaked in April 2026. Signals point to authentic.
• The internet decided the lesson was "emotional framing". That is the wrong lesson.
• The real lesson is ten structural patterns: modular named tags, direct imperatives, repeated critical rules, positive framing, explicit output format, tool discovery, scoped safety, default to helpful, possessive framing, and describing the shape of the work rather than its content.
• Apply the structure, not the content, and your own prompts will get measurably better.