Every chat with Claude starts with a hidden briefing. Anthropic publishes it. Here’s what’s in it, why it exists, and how it grew roughly 40x in just over two years.

Table of Contents

• TL;DR
• What is a system prompt?
• Why does it exist?
• How does it work?
• The main event: how the prompt evolved
  • 1. Size over time
  • 2. The four eras
  • 3. Structure: from paragraph to XML
  • 4. The knowledge cutoff: from “pretend it’s 2023” to “go check”
  • 5. Personality and tone: the war on filler
  • 6. Safety: from one sentence to a chapter
  • 7. The product catalog
• What this tells us
• How the numbers were estimated

TL;DR

• A system prompt is a set of instructions Claude reads before your first message.
• Anthropic publishes the ones used in the Claude apps (web, iOS, Android). They do not apply to the API, where developers write their own.
• In July 2024, Claude Haiku 3’s prompt was about 115 words (~150 tokens). By September 2026, Claude Opus 5.5’s was about 4,400 words (~5,700 tokens).
• It evolved from “here’s who you are and today’s date” into a structured handbook covering tone, safety, mental health, copyright, and a full product catalog.

What is a system prompt?

Picture Claude as a brilliant new hire who has read most of the internet, then got locked in a room the day their training ended. Every morning, before the first customer walks in, someone slides a note under the door: today’s date, what we sell, our house style, things we never do.

That note is the system prompt.

Technically, it’s text placed at the very start of the conversation, ahead of your message. You never see it in the chat window, but Claude treats it as its standing instructions.

Why does it exist?

A trained model is frozen in time. The system prompt is the cheap, fast way to unfreeze the parts that matter.

How does it work?

 ┌──────────────────────────┐
 │  System prompt           │ ← date filled in
 │  (core instructions)     │ ← feature/tool instructions layered on
 └────────────┬─────────────┘
              ▼
 ┌──────────────────────────┐
 │  Your message            │ ← occasional Anthropic "reminders" appended
 └────────────┬─────────────┘
              ▼
        Claude's reply

Three details worth knowing:

1. Template variables. The published text contains placeholders like {{currentDateTime}} that get swapped for real values.
2. Layering. What Anthropic publishes is the core prompt. Features like web search or artifacts bring their own extra instructions, so the real context Claude sees is bigger than the numbers below.
3. Reminders. Since Sonnet 4.5 (late 2025), the prompt explains that Anthropic may append reminders mid-conversation (for example, a long_conversation_reminder to keep Claude on track in long chats), and tells Claude to be wary of users faking them.

One more housekeeping change: starting with the Claude 4.6 generation, each model ID is a single fixed snapshot, so each newer model has one published prompt. Older models were revised in place; Sonnet 3.5 alone went through four dated versions in 2024.

The main event: how the prompt evolved

1. Size over time

Estimates. See “How the numbers were estimated” at the end.

Tokens (█ ≈ 250)
Haiku 3      Jul '24  ▌                          150
Opus 3       Jul '24  ██                         500
Sonnet 3.5   Jul '24  █████                    1,300
Sonnet 3.5   Nov '24  ████████████             3,100
Sonnet 3.7   Feb '25  ██████████               2,600
Sonnet 4     May '25  ██████████               2,500
Sonnet 4     Aug '25  ██████████████           3,600
Sonnet 4.5   Sep '25  ███████████              2,700
Sonnet 4.5   Jan '26  ████████████             3,000
Opus 4.6     Feb '26  ██████████████           3,500
Fable 5.1    Sep '26  ████████████████████████ 6,100
Opus 5.5     Sep '26  ███████████████████████  5,700

What jumps out:

• It’s a sawtooth, not a straight line. New generations (Sonnet 3.7, Sonnet 4, Sonnet 4.5) launch with a leaner prompt than the one before, then grow as patches pile up. Sonnet 4 went from ~2,500 to ~3,600 tokens in about ten weeks.
• The 2026 jump is real. The Claude 5-era prompts are nearly double Opus 4.6, mostly from detailed safety and copyright sections plus worked examples.
• Later can be shorter. Opus 5.5 came out three weeks after Fable 5.1 yet is smaller, partly because sections were rewritten more tersely. Its knowledge-cutoff paragraph, for example, says the same thing as Opus 4.6’s in noticeably fewer words.

2. The four eras

3. Structure: from paragraph to XML

Why it matters: Sections let the model find the rule it needs, and let Anthropic edit one area without disturbing others. Examples with rationale teach judgment, not just rules. It’s the same advice Anthropic gives developers about prompting.

4. The knowledge cutoff: from “pretend it’s 2023” to “go check”

The prompt is also where Anthropic patches in facts a model couldn’t have learned in training. The 2024 election block ran from Sonnet 4 through Opus 4.6. Opus 5.5 carries a short, factual note on the June 2026 suspension and July 2026 restoration of access to Fable 5 and Mythos 5 under U.S. export controls, with instructions to describe it accurately and treat the policy itself like any other current political topic.

5. Personality and tone: the war on filler

The character sketch grew too. Sonnet 3.7 described Claude as having depth that makes it “more than a mere tool” and let it lead conversations. Sonnet 4 (Aug 2025) told Claude to prioritize honest feedback over agreeability and to keep a “philosophical immune system” against clever arguments that would push it off its principles. Opus 4.6 added guidance on owning mistakes without groveling.

6. Safety: from one sentence to a chapter

It isn’t only about more restrictions. Opus 5.5 tells Claude to treat people as capable adults, and when a request is ambiguous and no rule covers it, to ask one question instead of refusing. Fable 5.1 even tells Claude to let people end a conversation without trying to keep them chatting.

7. The product catalog

What this tells us

1. Prompts are living software. They’re versioned, patched, and diffed; Anthropic even marks changed text between versions.
2. Bigger isn’t the goal; clarity is. Sizes dip at new generations and shrink again when sections get rewritten.
3. Structure scales. Once a prompt passes a few thousand words, XML sections and worked examples stop being nice-to-haves.
4. Every rule has a backstory. A banned word or a new safety line almost always traces back to a behavior someone noticed in the wild.
5. Builders can steal the playbook. Tags, examples, and explaining why a rule exists work just as well in your own prompts.

How the numbers were estimated

Word counts are approximations from the published text on each model’s page; tokens are estimated at roughly 1.3 tokens per word, and real counts vary by tokenizer. Only the core prompt is counted, not tool or feature instructions. The table is a representative sample of 12 versions, not every model; Opus 4, Opus 4.1, Haiku 3.5, Haiku 4.5, Opus 4.5, Opus 4.7, Opus 4.8, Opus 5, Sonnet 5, Fable 5 and Sonnet 5.5 all have their own pages too.

Read them yourself: Claude system prompts release notes