Claude 4.7 system prompt changes: what brands need to know about AI answer visibility
By be-recommended Team
TL;DR
Anthropic shipped Claude Opus 4.7 with a new system prompt, and Simon Willison published a side-by-side diff on April 18, 2026. For brand-visibility teams this matters because Claude's system prompt directly controls which sources it cites, how it formulates recommendations, and what it treats as trustworthy. Each prompt revision effectively rewrites the rules of how your brand surfaces in AI answers. The practical response is not to chase specific prompt lines, but to harden the signals every Claude version rewards: entity clarity, third-party citations, clean markdown sources, and documented first-hand experience.
Key Takeaways
- Claude's system prompt is public and versioned. Anthropic has published updated prompts with every major release since Claude 3 in July 2024.
- A prompt change is a ranking change. Behaviour around citations, hedging, self-disclosure, and "recommendation" framing shifts between versions.
- You can read the diff. Willison's post links a GitHub timeline of every Claude system prompt to date.
- Brands win by optimizing for signals that persist across versions, not by reverse-engineering line 42 of the current prompt.
What actually changed
On April 18, Simon Willison published Changes in the system prompt between Claude Opus 4.6 and 4.7 on his blog. The post walks through Anthropic's newly shipped instructions and points to a companion GitHub repository that tracks the full commit history of every Claude system prompt. Willison's analysis landed at number 11 on Hacker News within 18 hours with 262 points and 152 comments, which is a strong signal that the AI/dev community is paying attention to exactly this layer of the stack.
Anthropic publishing these prompts is unusual among frontier labs, and it gives brand-visibility teams something we rarely get with closed AI systems: a readable specification of how the model is told to behave. If you have ever wondered why Claude names one vendor over another, or refuses to endorse a specific tool, the answer often lives in a handful of sentences in the system prompt.
Why system prompt changes move the GEO needle
Claude is one of the three AI search surfaces that matter for brand discovery alongside ChatGPT and Perplexity. When its system prompt changes, three things shift for brands.
First, citation preference. The prompt instructs Claude on when to cite, how many sources to pull from, and how to frame those sources. Tighten the "cite authoritative sources" clause and recently-indexed Reddit threads lose airtime to established publications. Loosen it and the opposite happens. You see this ripple in AI answers within days of a release.
Second, recommendation language. Claude 4.x is consistently cautious about direct product endorsements and prefers comparative framings like "tools in this category include X, Y, and Z." Small wording changes in the prompt can expand or contract that list, which is exactly where your brand either shows up or gets skipped.
Third, self-disclosure and hedging. Prompts that encourage the model to flag uncertainty or to ask clarifying questions reduce the number of answers that name a specific brand at all. More questions, fewer recommendations.
None of those three levers is ever revealed in a marketing changelog. You only see them by reading the prompt diff.
What brands should do this week
The temptation is to sprint after specific lines in the 4.7 prompt. That is noise. Signal comes from hardening what every recent Claude version rewards:
- Clean entity pages. Your brand, products, and people should have structured Organization and Product schema with unambiguous names and descriptions. Claude grounds its recommendations against entity graphs before it opens your marketing copy.
- Third-party citations. Be named by publications Claude already trusts. Independent reviews, category analyses, and listings on comparison sites carry more weight than your own blog.
- Markdown-first sources. AI crawlers parse markdown natively. If you publish long-form content, expose a
.mdmirror for every key page. We covered the nuance in our llms.txt adoption post. - Documented first-hand experience. Prompts keep strengthening the E-E-A-T bias. Case studies with numbers and dates outperform generic explainer content.
- Track your share of voice in Claude specifically. Claude's answers differ materially from ChatGPT and Perplexity. Measure each surface separately.
The brands that get recommended across version 4.6, 4.7, and whatever comes next are the ones whose signals survive any reasonable rewrite of the system prompt.
FAQ
Where can I read the Claude 4.7 system prompt diff?
Simon Willison's post from April 18, 2026 is the most readable entry point: simonwillison.net/2026/Apr/18/opus-system-prompt. He links a GitHub repository that reconstructs the full prompt history as git commits.
Does Anthropic document every system prompt change?
Anthropic publishes the prompts themselves on their official Claude documentation, but not a full changelog of prompt intent. The community, led by Willison, has been doing the comparative work since Claude 3 in July 2024.
Should I rewrite my content for every Claude release?
No. Versions ship every few months and the specific wording will keep moving. Optimize for the signals that are stable across versions: entity structure, third-party citations, markdown mirrors, and documented experience.
How do I measure if Claude is recommending my brand?
Run consistent query sets against Claude weekly and track mentions, positions, and citation sources. Be Recommended automates this across Claude, ChatGPT, and Perplexity so you see when a system prompt change moves your visibility.
Sources
Stay ahead of the AI curve
Get weekly insights on AI visibility, search optimization, and brand strategy — straight to your inbox.