Claude System Prompt 2026: Anthropic Now Publishes the Full Internal Instructions
Last updated: August 2026
Anthropic has published an official, versioned documentation page that exposes the complete system prompts behind Claude's chat apps — the web interface at claude.ai plus the iOS and Android apps — for fifteen models stretching from Claude Haiku 3.5 to Claude Opus 5 and Claude Fable 5. Every prompt update is dated, and changes between versions are marked in bold, so anyone can now diff how Anthropic instructs its models over time. It is the answer to a question developers have asked since the first Claude release: what does the factory tell the AI before your message ever arrives?
What exactly is on the page?
The page lives inside Claude's official platform documentation under "System Prompts", and it covers the consumer chat apps only — not the API. Anthropic explains in the introduction that the system prompt feeds Claude up-to-date context such as the current date, encourages certain behaviors like always rendering code snippets in Markdown, and is periodically updated to improve responses. Those app-prompt updates, the docs note, do not apply to the Claude API at all.
Two design choices make the page genuinely useful rather than a one-off transparency gesture:
- One section per model. There are currently fifteen of them: Opus 5, Fable 5, Opus 4.8, Opus 4.7, Sonnet 4.6, Opus 4.6, Opus 4.5, Haiku 4.5, Sonnet 4.5, Opus 4.1, Opus 4, Sonnet 4, Sonnet 3.7, Sonnet 3.5, and Haiku 3.5.
- Dated entries with bolded diffs. Older generations list multiple dated prompt entries showing what changed and when. Starting with the Claude 4.6 generation, the docs explain, each model ID became a "single fixed snapshot" — so those models carry exactly one documented entry each.
| Generation | Documented models | Documentation style |
|---|---|---|
| Latest 2026 generation | Opus 5, Fable 5 | One fixed snapshot per model |
| 4.6 – 4.8 generation | Opus 4.8, Opus 4.7, Sonnet 4.6, Opus 4.6 | Single fixed snapshot |
| 4.5 generation | Opus 4.5, Haiku 4.5, Sonnet 4.5 | Dated entries with diffs |
| Older generations | Opus 4.1, Opus 4, Sonnet 4, Sonnet 3.7, Sonnet 3.5, Haiku 3.5 | Dated entries with diffs |
Source: Anthropic — Claude Opus 5 announcement
The detail everyone is talking about: Fable 5 requests being rerouted to Opus 5
Among the dated entries, one passage in the July 24, 2026 Claude Opus 5 prompt caught developers' attention. The prompt opens with internal tags — <claude_behavior> and then <product_information> — and tells the model plainly: "The currently selected version of Claude is Claude Opus 5. Claude Opus 5 is a powerful model for complex challenges."
The next part is the remarkable one. The prompt informs Opus 5 that a user's request may have been intended for Claude Fable 5 but redirected to Opus 5 instead due to a safeguards routing mechanism. It even acknowledges, in Anthropic's own words inside the prompt, that "the user may be confused about this situation (it's very recent!)" — and instructs Claude to answer questions about it by citing a quote from Anthropic's blog post on the subject.
That is an unusual degree of openness. Companies reroute traffic between models all the time and almost never document it. Here, the operational decision is baked into a now-public prompt, dated, and diffable. Developers who track these documents also note how much the prompts have grown: early system prompts ran a few hundred words, while the latest ones exceed three thousand — accreting rules the way building codes accrete after every incident.
Source: Anthropic — Claude Sonnet 5 announcement
How to track changes between versions yourself
The page's real value is not a single reading — it is the diff. Simon Willison, the developer who has tracked Claude prompts most consistently in public, keeps copies of them in a GitHub repository with a commit-style history, and has published detailed breakdowns of what changed between models such as Opus 4.6 and Opus 4.7. In practice this means that when Anthropic tweaks how Claude behaves in its apps, the change can be spotted within hours instead of surfacing as scattered user complaints.
A practical workflow: open the section for the model you use, read the paragraphs that matter to you — how it handles code, how it deals with ambiguity, what it says about its own capabilities — and skim the bolded changes between dated entries. If Claude suddenly feels different this month than last month, the reason is very likely written there.
Why this matters for students and professionals worldwide
If you are studying, building, or simply trying to get more reliable output from AI tools, this document has three concrete uses:
- A masterclass in prompt engineering, free. Anthropic's prompts are production-grade examples of instruction writing: explicit structure, ordered priorities, and honest handling of edge cases. Anyone building a chatbot, a research assistant, or a customer-support agent can study how a frontier lab organizes its instructions — then apply the same discipline to their own prompts, alongside practical starting points like our ready-made Gemini prompt templates for students.
- You learn what the model actually knows about you. The prompt reveals exactly what is injected before your message: the current date, a description of the interface, behavioral rules. That removes much of the mystery and makes your expectations more accurate.
- You can read model news critically. When the next model ships — whether it is a Claude release like Opus 5 and Sonnet 5 or an open-weights release like Qwen 3.8 27B — you can check documented behavior instead of relying on secondhand impressions.
Students using Truescho's AI study tools — flashcards, quizzes, and summaries — can treat these documents as a behind-the-scenes look at how well-built study assistants are instructed to behave: clear communication rules, explicit limits, and structured output formats.
A quick comparison: what the page covers and what it does not
| Question | Answer from the official docs |
|---|---|
| claude.ai and mobile app prompts? | Yes — complete, dated |
| API prompts? | No — those are the ones you write yourself |
| Claude Code prompts? | Not published on this page (developers note they are easy to extract with a logging proxy) |
| Changes between versions? | Bolded, with a date per update |
| Language of the prompts? | Published in English |
The historical contrast matters too. Before this page existed, public knowledge of Claude's prompts was assembled from leaks and independent analysis — Willison's archive being the best known. Now the official source precedes the analysts, and the community builds diffing tools on top of it rather than guessing.
Honest limitations to keep in mind
This is not "all the secrets of AI revealed." The page does not include API prompts (which are yours, not Anthropic's), does not include Claude Code's prompt, and a dated changelog is not a live feed — interim tweaks between dated entries would not appear instantly. Older generations list several dated entries while newer generations are a single frozen snapshot, so always check the last-updated date for the section you are reading. And a system prompt is strong guidance, not the only guardrail: providers also run external classifiers and filters that are not published anywhere.
Frequently Asked Questions
Does Claude's system prompt apply to API calls?
No. Anthropic's documentation states these system prompts are for the claude.ai web interface and the iOS and Android apps only. When you call the API, you write your own system prompt, and the app prompts are not added to your requests.
Can I read the full Claude Opus 5 system prompt right now?
Yes. The official page includes the complete text under the Opus 5 section, with the latest dated entry from July 24, 2026, and any changes from earlier versions marked in bold.
Do these prompts change over time?
Yes — and that is the point of the page. Anthropic periodically updates the app prompts to improve responses, and each update is added with its date and a bolded diff, while models from the 4.6 generation onward carry a single fixed snapshot because their model IDs no longer change.
I am not a developer. Why should I care?
Understanding how assistants are instructed makes you better at using every AI tool. Seeing how professionals structure instructions helps you write clearer requests for summarizing, translating, or planning your studies — and gets you more consistent results.
Why is an official publication better than leaks?
A leak is a context-free snapshot that may be outdated or altered. An official document is dated, comparable across versions, and carries accountability for accuracy — which is exactly what developers need to build reliable tracking tools on top of it.
Your next step
Open the official page from the sources below and jump straight to the section for the model you use — ten minutes of reading there is a miniature course in professional prompt writing. Then put it to work: try Truescho's AI study tools for your coursework, or read our Claude vs ChatGPT vs Gemini comparison to pick the right model for each task. Transparency has become a competitive feature — and you are the one who benefits.
Sources
- Anthropic — System Prompts (official documentation) — the official system-prompt page with the dated changelog
- Anthropic — Claude Opus 5 announcement — official announcement for the model referenced above
- Hacker News discussion — 265+ technical comments within the first day
- Simon Willison — Claude Opus prompt change analysis — trusted independent analysis of version-to-version changes
- simonw/research on GitHub — community archive tracking prompt changes