Claude's system prompt grew ninefold in two years. Version yours like code.

Anthropic's system-prompt release notes hit the top of Hacker News this week, collecting more than 550 points and 230 comments inside a day ^[1]. What drew the crowd was not any single instruction but the growth curve: by the count of a dev.to writeup, Claude Opus 3's prompt, dated July 12, 2024, runs 358 words, while Claude Opus 5's, dated July 24, 2026, runs 3,235 ^[2][3].

That is roughly nine times the length ^[4], about 2,900 words added at an average pace near 120 words a month ^[5]. The page covers claude.ai and the mobile apps, where a system prompt supplies up-to-date information such as the current date at the start of every conversation ^[6]; Anthropic states the updates do not apply to the Claude API ^[7]. Since the Claude 4.6 generation, each model ID is a single fixed snapshot, so each model gets exactly one changelog entry ^[8]. Simon Willison copied the page into a git repository holding 29 revisions across 17 models, each commit carrying the date from the source document, which makes a diff between any two versions of Claude's personality a one-line command ^[9]. That works out to 1.7 revisions per model to date ^[10], a figure the snapshot policy now caps at one going forward ^[8].

The Opus 5 text is sectioned like a spec: product information, Claude behavior, refusal handling, legal and financial advice, tone and formatting, user wellbeing ^[11]. Part of it is a product catalog, naming Claude Fable 5, Opus 5 as the currently selected model, Sonnet 5 and Haiku 4.5 along with their API strings ^[12], plus the tier above, Claude Mythos Preview, described as unavailable publicly and in use by a small number of trusted organizations under Anthropic's Project Glasswing ^[13].

Part of it is a news wire. The prompt records that Fable 5 and Mythos 5 were first released on June 9, 2026, that Anthropic suspended access on June 12 to comply with U.S. Department of Commerce export controls, that the Department lifted those controls on June 30, and that access was restored on July 1 ^[14], a 19-day gap ^[15]. It then tells the model these events postdate its training data, so it knows them only from the notice ^[16]. The writeup's reading is that a lab needing the model to know something new edits the prompt instead of fine-tuning ^[17], and the file supports it.

The same file also does routing and policy work. It says users can switch models mid-conversation, so earlier messages identifying as a different model or cutoff may still be accurate ^[18]; a Hacker News commenter noted Opus 5 is told a request intended for Fable 5 may have been redirected to it by a safeguard, and should be answered as if it were Fable 5 ^[19]. Child-safety rules are explicit about not supplying unstated assumptions that make a request directed at a minor look safer than written ^[20]. There is a roster of classifier-driven reminders Anthropic can inject mid-conversation, including image_reminder, cyber_warning, ethics_reminder, ip_reminder and long_conversation_reminder ^[21]. Even the cutoff is behavioral: reliable knowledge ends at the end of May 2026 ^[22].

The transferable part is the shape, not the content. A lab that could retrain instead writes post-cutoff facts, product tiers, routing behavior and legal caveats into text, dates each entry, and publishes it ^{[9][12][14][18]}. Most production prompts are still an undated string constant with no record of which behavior was added to fix which incident.

Watch whether the one-entry-per-snapshot discipline holds as more models ship, because that is what keeps the diffs readable ^[8][9]. Watch the injected-reminder list ^[21], which is a second, dynamic layer the published text only documents. And measure your own growth rate against 120 words a month ^[5].