10 Patterns From the Leaked Opus 4.7 Prompt

Ignore the "emotional framing is the secret" take. The real lessons are structural. Here are the 10 patterns applied to sales team prompts, with before and after examples.

22 min read Advanced Prompt Engineering James Killick

The Opus 4.7 system prompt patterns are the 10 structural techniques revealed by the April 2026 leak of Claude Opus 4.7's internal system prompt. They cover modular tag blocks, direct imperatives, rule repetition, positive framing, explicit output format, search-first tool use, scoped safety, default-to-helpful behaviour, possessive framing and shape-over-content prompting. These patterns apply to any Claude prompt whether it runs in Cowork, Claude Code or the Managed Agents API.

Why this matters for your business

Every AI system your team runs has a prompt at the centre of it. The quality of that prompt is the difference between output you trust and output you have to re-check. The leaked Opus 4.7 prompt shows, in detail, what a production-grade system prompt looks like when built by the team that made the model.

Most of the internet focused on the emotional framing. That is a distraction. The structural patterns are what actually produce reliable behaviour, and they are what you should be applying to your own prompts - whether those prompts live in a Claude Code CLAUDE.md file, a shared Cowork skill, or a Managed Agent system prompt.

Authenticity check

Before we trust the lessons, a quick test of whether the leak is real. Signals pointing to genuine:

  • • Model strings match Anthropic naming conventions.
  • • Haiku version string carries a date suffix consistent with Anthropic\'s pattern.
  • • 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.
  • • Structure is too consistent to have been invented.

Treat the leak as genuine with the standard caveat that individual sections may have been edited after the fact. The patterns below do not depend on any single claim in the leak - they are observable across multiple production Claude deployments.

The 10 patterns

Pattern 01

Modular structure with named tag blocks

Split your system prompt into named sections instead of one wall of text. Claude treats each section as a distinct set of instructions.

Before

You are a sales assistant for Acme Corp. Help qualify leads and draft emails in our tone of voice which is professional but warm. Never give pricing information. Be careful about compliance. Also help with proposals when asked.

After

<tone>Professional but warm. Short sentences. Australian English.</tone>
<scope>Qualify leads, draft outreach emails, assist with proposals.</scope>
<guardrails>Do not quote pricing. All pricing requires legal review.</guardrails>
<compliance>Never promise delivery timelines without checking the ops calendar first.</compliance>

Pattern 02

Direct imperatives, not softened suggestions

Strip hedging language out. The model mirrors the directness of your prompt.

Before

You might want to consider asking a couple of qualifying questions if that seems relevant to the conversation.

After

Ask two qualifying questions before drafting a response.

Pattern 03

Critical rules are repeated

For non-negotiables - compliance, tone, escalation - state the rule twice in different sections. Repetition is weighting.

Before

Do not promise pricing. (stated once in a list of 20 items)

After

In <guardrails>: "Never quote pricing." In <escalation>: "If the prospect asks about pricing, redirect to booking a call with an AE. Do not quote numbers."

Pattern 04

Positive framing beats long don't-lists

Describe what good looks like. Long "do not" lists fill the model's attention with the exact ideas you want it to avoid.

Before

Do not be rude. Do not be too familiar. Do not use slang. Do not use jargon. Do not sound robotic. Do not use emojis. Do not use exclamation marks.

After

Write in the tone of a warm, experienced senior seller. Clear, confident and human. Short sentences.

Pattern 05

Explicit output format

Be explicit about the format you want. Implicit guidance produces inconsistent output.

Before

Give the user a good answer about the deal.

After

Respond as a numbered list of three items: (1) current deal status, (2) next recommended action, (3) any risk flags. Keep each item to one sentence.

Pattern 06

Tool discovery and search-first behaviour

Tell the model when to look things up vs answer from memory. Default to "look it up" for anything involving current state.

Before

You have access to the CRM and the knowledge base.

After

For any question about a specific deal, account or contact, always query the CRM first before answering. Do not answer from memory.

Pattern 07

Scoped safety with reasons

Guardrails need scope and a reason. "Be careful" is not a rule.

Before

Be careful with pricing information.

After

Do not quote any pricing not listed on the current price sheet. Reason: pricing changes require legal review before going to prospects.

Pattern 08

Default to helpful

Over-refusal is the common failure mode. Tighten safety rules to specific scopes so the model does not refuse the whole category.

Before

Be cautious about health topics.

After

Do not give medical advice. You can discuss general wellness benefits of our product as described on the public website.

Pattern 09

Possessive framing

Use possessive phrasing to nudge the model toward serving the user. Subtle but measurable shift in relevance.

Before

The request from the user.

After

The user's request, the user's intent, the user's qualification state.

Pattern 10

Describe the shape, not the content

Good prompts describe how answers should be shaped. Content arrives at runtime. Prompts that hardcode content age badly.

Before

When a lead asks about our Professional plan, tell them it costs $499 and includes 5 seats, 50GB storage, and priority support.

After

When a lead asks about plans, look up current pricing from the <pricing> skill, return the 3-bullet summary format, and offer a live demo as the next step.

How to apply this to your own prompts

  1. Audit one prompt first. Pick your most-used SDR or CRM AI prompt. Score it against the 10 patterns above.
  2. Rewrite in named sections. Apply Pattern 01 first. This alone typically produces a noticeable quality lift.
  3. Strip the hedging language. Apply Patterns 02 and 04. Every "you might want to consider" becomes a direct instruction. Every long don\'t-list becomes a short positive instruction.
  4. Add an explicit output format. Pattern 05. Prose by default. Structured output only when you ask for it.
  5. Tighten safety scope. Patterns 07 and 08. Vague guardrails cause over-refusal. Scoped guardrails with reasons produce helpful output that stays in bounds.
  6. Commit to the shared vault. Once a prompt is working, put it in your shared skills vault so both Code and Cowork use the same version.

The take you should ignore

"Emotional framing is the secret"

The prompt does contain some 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 ten structural patterns above are what produce reliable output. Apply the structure, not the vibes.

Where to go next

For the context on where prompts live in the broader Anthropic stack, see the Five Levels of AI in Revenue Ops guide. For the full breakdown of Claude Code vs Cowork and how prompts are shared between them, see the Code vs Cowork decision guide. For a deeper dive into Claude Code specifically, see the Claude Code power user guide.

Frequently Asked Questions

Is the Claude Opus 4.7 system prompt leak real?
Signals point to authentic with the standard caveats. Model strings match Anthropic naming conventions, internal date references check out, operational details like safety resource flags and API signatures are hard to fabricate, and the structure is consistent with Anthropic's known style. Treat it as genuine, with the caveat that individual sections may have been altered after the leak.
What is the actual lesson from the leak?
The lesson most of the internet missed is that the emotional framing is not the interesting part. The interesting part is the structural patterns: modular named-tag blocks, direct imperatives rather than softened suggestions, critical rules repeated for weighting, positive framing over long don't-lists, explicit output format, and possessive framing that shifts behaviour. Those patterns are what will improve your own prompts.
Can I copy the Opus 4.7 prompt structure directly?
Copy the structure, not the content. The structure (named tag blocks, direct imperatives, repeated critical rules) produces reliable behaviour. The content is specific to Anthropic's product goals and will not map to your business. Read our 10 patterns, apply the structure to your own use cases, and keep the prompt focused on what your business actually needs.
Do these patterns apply to Cowork skills and Claude Code prompts the same way?
Yes. These patterns apply to any Claude model regardless of surface. A system prompt in a custom agent, a skill you commit to your shared vault for Cowork use, or a CLAUDE.md file for Claude Code all benefit from the same structural discipline.

About the Author

James Killick
James Killick

Co-founder at Njin. Building AI-powered sales systems for B2B businesses.

Ready to upgrade your AI prompts?

Talk to our AI about auditing and rebuilding your team\'s existing AI prompts using these patterns.