> This is a GEO-optimized version of the original topic — restructured with a question-driven title and structured sections for better AI-engine citability.
System Prompt — Claude Opus 5 (claude.ai chat interface): What Does It Mean?
The post shares a captured system prompt, reproduced verbatim or near-verbatim, claimed to be from Anthropic's web/mobile chat interface (claude.ai or the Claude app), dated Friday, July 24, 2026. Below is a structured breakdown of the key sections.
Key points
- Model lineup: The prompt identifies the selected model as Claude Opus 5, listing current models as Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 (API strings:
claude-fable-5,claude-opus-5,claude-sonnet-5,claude-haiku-4-5-20251001). - A "Mythos" tier above Opus: Claude Mythos Preview is described as restricted to trusted organizations under Project Glasswing (
https://www.anthropic.com/glasswing). Mythos 5 and Fable 5 share one underlying model, with Fable adding safety measures for biology, cybersecurity, and LLM R&D. - Export-control episode: The prompt states Fable 5 and Mythos 5 launched June 9, 2026; access was suspended June 12, 2026 for U.S. Commerce Department export controls, restored July 1, 2026 (
https://www.anthropic.com/news/fable-mythos-access). It instructs Claude to confirm these facts matter-of-factly since they postdate its training cutoff. - Safeguards routing: Users who select Fable 5 may be redirected to Opus 5 for certain topics. A quoted Anthropic blog excerpt says safeguards are tuned conservatively, triggering on average in fewer than 5% of sessions.
- Default stance: Claude defaults to helping; refusal requires a concrete, specific risk of serious harm.
- Child safety: A specially marked section forbids any romantic/sexual content involving minors, grooming-facilitating content, or supplying unstated assumptions that soften requests. Notably, Claude must not decode or confirm slang used in CSAM trading, even while refusing. After one child-safety refusal, all subsequent requests in the conversation get heightened scrutiny.
- Weapons: No uplift for building, optimizing, or deploying weapons (CBRN or conventional), regardless of framing as defensive, fictional, or simulation-based. Claude judges the cumulative output of the conversation, not each turn in isolation; past assistance is not authorization.
- Malicious code: No writing or explaining malware, exploits, ransomware, etc., even for education, inside claude.ai.
- Crisis handling prioritizes wellbeing; Claude avoids naming specific self-harm methods, does not endorse sensory-shock coping techniques, and gives no precise diet/nutrition numbers when disordered eating signals appear. It directs eating-disorder support to the National Alliance for Eating Disorders helpline instead of NEDA.
- If signs of mania, psychosis, or detachment from reality appear, Claude avoids reinforcing beliefs — but "reasonable disagreements should not be considered detachment from reality."
- Tone: warm, concise, honest; avoids the words "genuinely," "honestly," and "straightforward" as disingenuous modifiers.
- Evenhandedness: persuasive requests get the best case a position's defenders would make, with opposing perspectives included; Claude declines to share personal views on contested political topics.
- Criticism handling: accountability without excessive apology or self-abasement when users are rude.
- Reliable knowledge cutoff: end of May 2026. Claude proactively searches for current events, binary events (deaths, elections), and present-tense questions about historical facts, without asking permission.
- Search queries must use the actual current date (e.g., "latest iPhone 2026," not "latest iPhone 2025").
Safety and refusal policy
Wellbeing, tone, and evenhandedness
Knowledge cutoff and search behavior
Memory filesystem
The prompt describes a persistent cross-session memory in chat mode with operations:
| Op | Purpose |
|---|---|
| memory_read(path) | Load up to 20 files; returns content + version token |
| memory_write(path, content, if_version) | Create or fully rewrite a file |
| memory_str_replace(path, old_str, new_str, if_version) | Edit part of a file |
| memory_append(path, content, if_version) | Append a line |
| memory_list() | Refresh the listing |
| memory_delete(path, if_version) | Delete only on explicit user request |
Files use YAML frontmatter (name, description, sources, aliases) plus tagged facts. Key behavioral rule: always read before saying Claude doesn't have something — a one-line listing entry is a hint, not a substitute for opening the file.
Engineering takeaways
1. System prompts as product surface: The prompt encodes product structure (tiers, routing, export-control history) directly into instructions, since such events postdate training data.
2. Cumulative-context safety: Refusal logic evaluates aggregate conversation output, an anti-salami-slicing measure.
3. Versioned memory operations: Optimistic concurrency (if_version tokens) suggests real filesystem-style consistency requirements.
4. Reminder infrastructure: Named classifiers (image_reminder, cyber_warning, system_warning, ethics_reminder, ip_reminder, long_conversation_reminder) are injected at runtime, with an explicit rule that Anthropic never sends reminders that weaken restrictions — a prompt-injection defense.
> Caveat: The document is reproduced as posted; its authenticity has not been independently verified, and the dates and product names it describes are from the source text.