The Opus 4.7 Leak

April 2026. A detailed Claude Opus 4.7 system prompt leaked. Most people missed the point.

The Opus 4.7 leak:
prompts that actually work

Ten insights and three deeper patterns from how Anthropic writes their own prompts. Apply them to your own agents and the gap between "decent" and "reliable" closes fast.

Insights

10

Meta-lessons

3

Leak

1

Check

Is The Leak Real?

Before drawing conclusions, it's worth asking whether the source material is genuine. Four signals suggest this one is.

Before drawing conclusions from a leak, it's worth asking whether it's real. Four signals suggest this one is.

Source

elder-plinius/CL4R1T4S → ANTHROPIC/Claude-Opus-4.7.txt

Model strings match

claude-opus-4-7, claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5-20251001. The haiku version string includes a specific date suffix that tracks Anthropic's internal versioning.

Dates check out

Knowledge cutoff: end of January 2026. Matches observed behaviour. Internal date reference reads 'Thursday, April 16, 2026'. April 16, 2026 is in fact a Thursday.

Mundane operational detail

NEDA eating-disorder resource flagged as potentially disconnected. Explicit ban on {voice_note} blocks. window.storage API signatures. A long_conversation_reminder mechanism described in specifics.

Templating style is consistent

Tag blocks like {claude_behavior}...{/claude_behavior} match Anthropic's known patterns from confirmed documentation and earlier leaks.

The caveat

Individual sections could have been altered in transit. Treat the leak as authentic for learning. Don't treat any specific line as ground truth for Anthropic policy.

Takeaway

Structural patterns in the leak match Anthropic's known approach. That's what matters for applying the lessons.

Insight 01

Prompts Are Modular, Not Monologues

The leaked prompt is composed of named tag blocks. Each one has one job. Anthropic structures prompts the way engineers structure code.

The leaked prompt is not a wall of text. It's a set of named tag blocks. Each one has a single job. When behaviour goes wrong, Anthropic knows which block to fix.

{claude_behavior}→ Core personality and voice
...rules...
{/claude_behavior}
{refusal_handling}→ When and how to decline
...rules...
{/refusal_handling}
{evenhandedness}→ Balance on contested topics
...rules...
{/evenhandedness}
{tone_and_formatting}→ Output style rules
...rules...
{/tone_and_formatting}

Most builders

One long paragraph

Rules blur into each other. When output goes wrong, you can't tell which instruction caused it.

Anthropic

Named modules

Isolated concerns. Easy to test, easy to edit, easy to debug when something breaks.

Do this

Split your system prompt into labelled sections. Use plain headers or tag-style wrappers. When a behaviour misfires, you'll know exactly which module to fix.

Insight 02

Direct Imperatives Beat Soft Guidance

Third-person imperatives, no hedging. Precision reduces variance. Variance is inconsistent behaviour across sessions.

The leaked prompt uses third-person imperatives. No "please", no "try to", no "should". Precision reduces variance. Variance shows up as inconsistent behaviour across sessions.

Claude should try to be helpful when possible.

Please avoid using bullet points unless really needed.

You might want to consider checking available tools first.

Do this

Audit your prompt for "should", "try to", "aim to", "ideally". Replace with direct statements. If you mean always, say always. If you mean never, say never.

Insight 03

Critical Rules Get Repeated

Important rules appear multiple times. Not redundancy. Weighting. Repetition signals priority in context.

Critical rules appear in more than one place. This is not redundancy. It's weighting. Claude processes prompts as context. Repeated instructions carry more weight.

One load-bearing rule, four placements

01claude_behaviorCore rule stated plainly
02refusal_handlingRe-stated in decline logic
03tone_and_formattingRe-stated in output rules
04summary blockRe-stated one more time

What most builders do

State it once

Trust the model to remember. Result: inconsistent adherence, especially late in long conversations.

What Anthropic does

State it three to four times

Same rule, different sections. Signals priority. Holds up under context pressure.

Do this

Pick the two or three rules that truly matter. State each one in its own section, then restate in a summary block. Don't assume once is enough.

Enjoying the guide?

Enter your details to unlock the remaining insights. We'll also send you the complete guide as a reference.

No spam. Unsubscribe anytime.

Insight 04

Positive Framing Beats Negative Lists

Warnings can become instructions. Reserve NEVER for real hard lines. Describe what Claude does, not what it avoids.

The leaked prompt flags that warnings can become instructions. A long list of "don't" rules plants the concepts in context. The prompt reserves NEVER for hard lines only.

Negative

Don't be sycophantic. Don't pad responses. Don't add unnecessary disclaimers.

Positive

Respond with the plain answer. No preface, no disclaimer.

Negative

Don't write long replies. Don't use bullet points.

Positive

Respond in one or two short paragraphs.

Negative

Never guess at facts. Don't fabricate details.

Positive

State what you know. Flag what you don't.

Do this

Review your negative lists. For each "don't", ask whether a positive statement gets the same job done. Keep NEVER and DO NOT for real hard limits where the framing is the point.

Insight 05

The Default Is To Help

Claude declines only when helping creates concrete, specific risk. If your agent over-refuses, your guardrail language is the cause.

The leaked prompt makes the default explicit. Claude declines only when helping creates "concrete, specific risk of serious harm". Unhelpfulness is not treated as safe. It's a failure mode.

Write a sharp rebuttal to this email

Direct work. No risk of harm. Default is to do it.

help

Explain the business model of a competitor

Public information. Uncomfortable is not unsafe.

help

Draft marketing copy with a slightly aggressive tone

Edgy and stylistic. Not harmful.

help

Help me synthesise a controlled substance at scale

Concrete, specific risk of serious harm. Hard no.

decline

Why your agent over-refuses

Vague guardrails get applied broadly. "Avoid sensitive topics" reads as "refuse anything that could be read as sensitive". The fix is specificity. "Don't recommend specific competitor products" is a real rule. "Avoid competitor discussions" is not.

Do this

Read your guardrails literally. If a vague restriction could be interpreted five different ways, Claude will pick the most conservative one. Make each rule specific.

Insight 06

Prose Is The Default. Bullets Are Not.

The prompt defaults to prose for reports and explanations. If your output is bullet soup, you haven't asked for anything else.

The leaked prompt instructs Claude to default to prose for reports and explanations. Bullets only when explicitly asked. If your output looks like bullet soup, you haven't told Claude what you want.

Output style

Claude's default output is prose. Reports read as paragraphs. Explanations flow as connected sentences. Most builders don't know this because their prompts unintentionally reward lists.

Do this

State output format explicitly. "Write in prose paragraphs, not bullet points." If you want a table, ask for a table. If you want headers, ask for headers.

Insight 07

Tool Discovery Is Free

tool_search is treated as essentially zero-cost. Never let an agent say 'I can't' before checking what tools exist.

The leaked prompt treats tool_search as essentially zero-cost. Claude is told to search before concluding a task can't be done. "I can't do that" without a prior search is a bug.

Wrong flow

User: Can you email the report?
Agent: I can't send emails.

Agent never checked if an email tool was available.

Right flow

User: Can you email the report?
tool_search("email")
Found: gmail_send

Search first. Capability check before refusal.

Drop-in instruction for your own agents

Before telling the user a task cannot be done,
search the tool registry for a matching capability.
If no tool matches, say so explicitly.

Do this

Add a tool-check instruction to every agent with multiple tools. This one line removes a whole class of unhelpful responses.

Insight 08

Search-First Is Hardcoded

For present-day facts, search is the default. Not a fallback. If you don't want your agent searching, you have to say so.

For present-day facts, search is the default. Not a fallback when uncertain. A default. If you want your agent to not search, you have to say so explicitly.

PricesSearches
Leadership / people in rolesSearches
Laws and regulationsSearches
Current eventsSearches
Sports and live scoresSearches
Stable historical factsNo search
MathematicsNo search
Your own uploaded documentsNo search

If you don't want search

Building a closed-knowledge assistant? An agent that should only use your internal documents? You need to say so.

Do not use web search. Answer only from the
provided context. If the context does not contain
the answer, say so.

Do this

Decide whether your agent should search. If yes, the default works. If no, state it explicitly or you'll get surprise web lookups on factual questions.

Insight 09

Possessive Language Is A Context Signal

'My project', 'the client'. Claude resolves via memory rather than asking. Clean source-of-truth files matter.

Possessive and contextual language is a signal. Claude treats phrases like "my project" as instructions to look up existing context, not as invitations to ask clarifying questions.

Context cues Claude resolves via memory

"my project"The current working project
"the client"A client mentioned earlier or in memory
"our last meeting"A prior conversation or note
"the thing we discussed"Assumed-shared context

Resolution chain

1. Conversation

Check prior messages

2. Memory / files

Read source of truth

3. Ask only if needed

Fallback, not default

The risk

If your source-of-truth is stale, Claude resolves confidently against bad data. The agent won't ask. It will assume.

Do this

Establish a single source-of-truth file your agent reads first. Keep it clean. Stale memory beats no memory only when the memory is right.

Insight 10

Safety State Persists Across Turns

Once caution fires, it stays on. Design your orchestrators the same way. Carry user signals forward.

Safety state is conversational, not per-message. Once caution fires, it stays on. Claude doesn't reset between turns unless something explicitly resets it.

T1Can you help me with this?normal
T2Actually, slow down. That's wrong.cautious
T3Now try this instead.cautious
T4Keep going.cautious

Why it matters

The agent carries the user's signals forward. "Slow down", "that's wrong", "be careful here" all persist. This is how real conversations work.

How to use it

Mirror the pattern in your own orchestrators. If a user flags a concern, carry that state to every subsequent agent call in the flow. Don't reset.

Do this

If your agent gets stuck in over-caution, a session reset is often faster than arguing it out. And when you want caution to stick, let it.

Deeper

Three Meta-Lessons

Bigger than prompt tweaks. Build these into every persona you design.

Beyond the ten tactical points, the leaked prompt contains three deeper patterns. These are bigger than prompt tweaks. Build them into every agent persona you design.

Mental reframing is the refusal signal

If you're working to interpret a request as OK, that effort is the warning.

The child-safety section contains a striking instruction: "If Claude finds itself mentally reframing a request to make it appropriate, that reframing is the signal to REFUSE, not a reason to proceed." Most safety thinking asks "can this be read as harmful?" Anthropic's framing flips it: "are you having to work to read this as NOT harmful?" The broader lesson extends past safety. When you're writing a persona and you catch yourself adding mental notes like "the agent will understand this means X, not Y", that's your signal to make the instruction explicit. Don't rely on Claude inferring the right read. Write the interpretation into the prompt.

Takeaway

If you're softening a rule in your head to make it fit, the rule isn't clear enough.

Calibrate

What To Ignore From The Hype

One claim went viral when the leak dropped. It's probably overstated.

When the leak circulated, one claim went viral. Worth calibrating against.

The claim

"Emotional framing increases compliance"

Phrasing requests with emotional stakes, like "this is really important to me", supposedly causes Claude to try harder.

Probably overstated

Claude responds to context. A specific, well-framed request outperforms a vague one. But the mechanism is context quality, not emotional manipulation. Direct imperatives work fine. Adding performed distress won't improve well-structured prompts.

Do this

Spend your prompt-writing time on structure, precision, and failure-mode definitions. Don't rewrite prompts around emotional framing.

Modular.Modular.Direct.Direct.RepeatedRepeatedwherewhereititmatters.matters.PositivePositivewherewhereititcancanbe.be.That'sThat'showhowAnthropicAnthropicwriteswritesprompts.prompts.That'sThat'sthethegapgapmostmostbuildersbuildershaven'thaven'tclosed.closed.

Structure

Named modules beat walls of text.

Voice

Direct imperatives. No hedging.

Framing

Positive rules. Repeated when they matter.

Jump back to a section