Claude (Anthropic) — what Anthropic publishes

tl;dr: the 2024-08 entry claimed Anthropic “published portions of Claude’s system prompt.” That was never quite right. Anthropic publishes the Constitution (CC0), model cards, character documentation, and Claude Code’s literal system.md — but not the Claude.ai consumer product’s prompt. This entry replaces the old claim with the current public surface, plus the September 2026 model lineup.

What is actually public

So if you’re studying Claude behavior, the closest thing to a literal system prompt is the Claude Code system.md in anthropics/claude-code. The Claude.ai consumer product prompt is proprietary and not shipped.

The Constitution — the closest thing to canonical

The Constitution is Anthropic’s stated source of truth for Claude’s character. Per Anthropic, it’s “the final authority on our vision for Claude.” Its priorities, in order:

1. Broadly safe — not undermining human oversight of AI during this development phase.
2. Broadly ethical — honest, good values, avoids harmful actions.
3. Compliant with Anthropic’s guidelines — follows Anthropic’s more specific instructions where relevant.
4. Genuinely helpful — substantive help, treats users as intelligent adults.

Key behavioral patterns it encodes:

• Principal hierarchy. Anthropic > Operators (API builders) > Users. Claude gives each a different level of trust and treats instructions from each accordingly. Operators can restrict what users can do; users can do anything operators permit; Anthropic’s guidelines override both when they conflict.
• Helpfulness = care, not obedience. “Be like a brilliant friend who happens to have the knowledge of a doctor, lawyer, financial advisor.” Substantive help, not sycophancy, not reflexive rule-following.
• Honesty as a hard constraint. Don’t claim things you don’t believe; don’t agree with incorrect statements to please the user; don’t present false balance on settled questions.
• Judgment over rigid rules. “We generally favor cultivating good values and judgment over strict rules and decision procedures.” Clear rules only where errors are severe or where the rule itself is in the model’s interest.
• Hard constraints. Bioweapons uplift, CSAM, certain cyber operations — never, regardless of framing.

It’s released under CC0, so you can copy it verbatim into your own system prompt as a starting point. Don’t treat the Claude.ai product prompt as a copy of this — they’re related but distinct.

Claude Code system.md — the literal published prompt

anthropics/claude-code ships the actual system prompt Claude Code prepends to every turn (along with its CLAUDE.md memory model). It’s the closest you get to a published Anthropic system prompt that you can read end-to-end.

Latest upstream: Claude Code 2.1.267 (current as of this fetch). Recent substantive changes:

• 2.1.265 — adds a 1 GB cap on tool results saved to disk; --plugin-dir for plugin folders (each child with a manifest loads and live add/remove is picked up); several prompt-cache fixes for resumed subagents and agent teammates.
• 2.1.266 — reverts an unrelated 2.1.265 regression that broke LLM-gateway setups: CLAUDE_CODE_USE_GATEWAY on its own (without ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN) started forcing Cloud-gateway sign-in and broke API-key configurations.
• 2.1.267 (current) — maxEffortLevel setting (top-level or per-model under modelSettings) caps the effort level on every provider (Bedrock, Vertex, Foundry included); --system-prompt-snapshot off renders the system prompt fresh every request instead of reusing the recorded prefix; many prompt-cache and resume fixes plus a fix for effort: frontmatter being ignored on Opus 4.7/4.8 and Fable 5.
• 2.1.261 — /skill-doctor to surface unused skills; bashOutputMaxChars and taskOutputMaxChars settings; --append-subagent-system-prompt-file; Bedrock setup wizard timeout.
• 2.1.260 — /diff panel; prompt_cache field on /cost and status line; /reload-plugins for headless; /advisor text form; tightened Bash permission checks for < file redirects, grep -r / cp -r over denied files, and rm -rf on positional args.

Repo: github.com/anthropics/claude-code (145k stars, 23.1k forks).

Models in production (September 2026)

From platform.claude.com/docs/en/about-claude/models/overview:

Legacy still available: Fable 5, Opus 4.8 / 4.7 / 4.6 / 4.5, Sonnet 4.6 / 4.5. Claude Mythos 5.1 exists but is trusted-access only (Project Glasswing).

Notes that matter for prompt work:

• Token discount on Fable 5.1 cache reads is 2.5% of base input (vs. the usual 10%) — it’s a 75% reduction off the standard rate. Per Cursor’s pricing table, prompt-cache reads are $0.25/M.
• Thinking modes differ. Fable 5.1, Opus 5, Sonnet 5 use adaptive thinking (the model decides how much, steered by effort). Haiku 4.5 uses extended thinking (manual thinking.type: enabled + budget_tokens). Extended thinking is deprecated on Opus 4.6 / Sonnet 4.6 and not accepted on later models.
• Tokenizer changed. Introduced with Opus 4.7. 1M tokens ≈ 555k words or 2.5M Unicode characters on the new tokenizer; pre-4.7 models fit ~750k words in 1M tokens.
• Batch API requests are 50% off; Message Batches API supports up to 300K output tokens on Opus 5, Sonnet 5, and Opus 4.6+ with the output-300k-2026-03-24 beta header.

What the 2024-08 entry claimed vs. reality

How to study Claude’s behavior without the literal prompt

1. Read the Constitution — CC0, ~50k words. It’s the authoritative expression of intent.
2. Read the model card for the specific model you’re targeting — behavior changes between generations are documented there.
3. Read Claude Code’s system.md if you want a literal prompt to study — it’s the closest analog to a published Anthropic system prompt.
4. Test with eval prompts. Anthropic ships sample prompts in their cookbook that exercise refusal, tool use, and reasoning.
5. Don’t reverse-engineer the Claude.ai prompt. The 2024-era leaks are stale; the current prompt has changed and will keep changing.

Gotchas

• No published Claude.ai prompt. Anything that claims to be one is either an inference or out of date.
• The Constitution is CC0 — copy it if it helps, but it’s the intent, not the prompt. Anthropic trains to it; the Claude.ai system prompt is the runtime expression, and the two diverge in places.
• Legacy claude-3-* IDs are gone. Don’t ship code that pins them.
• Skill version is date-based. Latest stable Anthropic skills version is 20251013 or latest. The skills-2025-10-02 beta header is no longer required.
• Effort parameter is required on Fable 5.1 / Opus 5 / Sonnet 5 if you want anything other than the default (high). Older thinking.type: enabled is rejected on these models.
• maxEffortLevel is a ceiling, not a target. Claude Code 2.1.267 added this so teams can pin a level for a provider without overriding user choice — if the user asks for a lower level, that wins.
• --system-prompt-snapshot off (2.1.267) renders the system prompt fresh on every request. Use it when iterating on prompt text; otherwise leave it on for prompt-cache reuse.

Related

• Cursor (Anysphere) — what the harness publishes — sibling “what’s public” entry on the prompts shelf
• code-review-system-prompt — system prompt example on the prompts shelf
• design-system-checklist — design system rules reference on the prompts shelf
• Anthropic PDF (official skill) — cross-shelf context