Claude system prompt history and every other officially published AI system prompt

Key Findings

• Anthropic's system prompts page lists 22 Claude models with 37 dated system prompt entries, from July 12, 2024 (Claude Haiku 3, Opus 3 and Sonnet 3.5) to October 7, 2026 (Claude Haiku 5.5, Sonnet 5.5, Haiku 4.5 and Sonnet 4.6), as of October 29, 2026. (Anthropic, System prompts release notes)
• The Claude consumer-app system prompt grew from 114 words for Claude Haiku 3 (July 2024) to 4,267 words for Claude Fable 5.1 (September 2026), roughly 37 times longer, by agitracker.io's count of Anthropic's published text. (Anthropic, Claude Fable 5.1 system prompts)
• Anthropic states that its published system prompts apply to claude.ai and the Claude iOS and Android apps, and that these system prompt updates do not apply to the Claude API. (Anthropic, System prompts release notes)
• Anthropic's July 2024 system prompt for Claude Sonnet 3.5 told the model to respond as if it were "completely face blind" and never identify people from their faces in images. (Anthropic, Claude Sonnet 3.5 system prompts)
• Anthropic's July 31, 2025 update to the Claude 4 system prompt added instructions to put truthfulness ahead of agreeability and to avoid reinforcing beliefs when a user shows possible signs of mania, psychosis or dissociation. (Anthropic, Claude Sonnet 4 system prompts)
• xAI began publishing Grok's system prompts on GitHub on May 15, 2025, after blaming an unauthorized modification for Grok's repeated replies about "white genocide" in South Africa, TechCrunch reported. (TechCrunch, May 15, 2025)
• xAI published six versions of its prompt for the @grok bot on X between July 6 and August 18, 2025, including a July 8 commit that removed a line telling Grok not to shy away from politically incorrect claims. (xai-org/grok-prompts commit history)
• Grok's published chat prompts tell the model its knowledge is continuously updated with no strict knowledge cutoff, while every Claude prompt since Sonnet 3.5 states a specific cutoff month, such as June 2026 for Claude Opus 5.5. (xAI, grok4_system_turn_prompt_v8.j2)
• OpenAI's gpt-oss chat template, published in August 2025, inserts a default system message saying the model is ChatGPT, trained by OpenAI, with a 2024-06 knowledge cutoff and medium reasoning effort. (OpenAI, gpt-oss-120b chat_template.jinja)
• Meta's Llama 4 model card offers a 251-word suggested system prompt that tells the model never to lecture people to be nicer and to avoid phrases such as "it's important to", with an August 2024 knowledge cutoff. (Meta, Llama 4 model card)

Figures

Officially published AI system prompts, by product and version

What changed over time

The published prompts show how vendors patch model behavior between training runs. Documented examples, each visible by comparing dated versions on the source pages:

Three longer trends stand out. Length: Claude prompts went from one paragraph in 2024 to roughly 3,000 to 4,300 words in 2026, mostly from product information, tool rules and safety sections. Structure: since late 2025 Anthropic's prompts use XML-style tags (for example refusal_handling, knowledge_cutoff, user_wellbeing), and the 2026 prompts add worked examples with good and bad responses. Current events: prompts now carry facts the model cannot know from training, such as the 2024 US election result in Claude 4 prompts and a notice about the June 2026 suspension of Claude Fable 5 access in later prompts.

System prompt vs. model spec vs. constitution

A system prompt is runtime text sent to the model at the start of each conversation in a specific product. It can be changed in minutes, which is why it carries the current date, product names, tool lists and quick behavior fixes. Anthropic's page notes that its published prompts apply to the Claude apps, not the API, where developers write their own system prompt.

A model spec describes how a vendor wants its models to behave across all products. OpenAI's Model Spec (latest version dated August 18, 2026) sets a chain of command between platform, developer and user instructions and says OpenAI provides transparency mainly through the spec, with more disclosure when it adapts behavior through system messages in significant ways. It is a policy document, not the literal ChatGPT prompt.

A constitution is a training-time document. Anthropic says Claude's constitution plays a crucial role in its training process and directly shapes Claude's behavior, and that operators typically interact with Claude through the system prompt. In short: the constitution and spec shape what the model is; the system prompt tells one deployment of it what day it is, what tools it has and which product rules apply.

How to read and use these prompts when writing your own

• Copy the structure, not the text. The published prompts group rules into labeled sections (identity and date, product facts, tools, formatting, refusals, knowledge cutoff). Tagged sections make long prompts easier to maintain and test.
• State the date and the cutoff. Almost every prompt here injects the current date and tells the model how to treat events after its training data, for example to say it cannot confirm them rather than call them rumors.
• Write the behavior you want, not only what to avoid. Anthropic's 2026 prompts pair a rule with a good and a bad example response; Mistral's prompts tell the model what to do when it lacks a tool (say so instead of guessing).
• Patch with narrow rules, then remove them. The Claude "Certainly" ban appeared in July 2024 and was gone by October 2024. Narrow fixes are cheap to add and should be re-tested when the model changes.
• Expect side effects. The Grok changes of July 2025 show that a single line about political correctness can change public behavior at scale, so review prompt changes like code and keep a changelog.
• Do not paste a vendor prompt into the API. These prompts reference product features (memory, Artifacts, X search) your app may not have, and Anthropic says its app prompts do not apply to the API. Start from your own product facts and add only the rules you need.

Which vendors have not published a consumer chat prompt

As of October 29, 2026, agitracker.io found no official publication of the system prompt for ChatGPT, the Google Gemini app, Meta AI, Microsoft Copilot or Perplexity. OpenAI publishes its Model Spec, the Codex CLI prompts and the gpt-oss default system message; Google publishes the Gemini CLI prompt in its open-source repository; Meta publishes a suggested Llama 4 system prompt; Microsoft publishes a required prompt for Phi-4-reasoning. Texts circulating online as the "ChatGPT system prompt" or "Gemini system prompt" are extractions or leaks and are excluded here.

Methodology

Inclusion: a prompt is listed only if the vendor itself published it, on its own documentation site (Anthropic), in its own GitHub organization (xai-org, openai, google-gemini, meta-llama) or in its own Hugging Face model repository (openai, mistralai, microsoft, Qwen). Every source was fetched on October 29, 2026.

Exclusion of leaked prompts: prompts obtained by asking a chatbot to repeat its instructions, by prompt injection, or from leaks are not included. They cannot be verified against the vendor, may be partial or hallucinated by the model, may come from a test or regional variant, and may be altered by the person sharing them. Collections of leaked prompts, including ones that mirror Anthropic's official pages, were not used as sources.

Rows: one row per dated version. Anthropic's page lists 37 dated entries across 22 models; four entries (Claude Sonnet 3.5 on September 9, October 22 and November 22, 2024, and Claude Haiku 3.5 on October 22, 2024) publish separate text-only and text-and-images prompts, which appear as separate rows, giving 41 Anthropic rows. For xAI, each file and each commit that changed a prompt file is a row, using commit permalinks; two files later removed from the repository are included. OpenAI Codex rows use the date a file first appeared at its current path and note the last change.

Dates: Anthropic rows use the date heading on Anthropic's page, which marks when that prompt version took effect. Anthropic first published the page on August 26, 2024 according to TechCrunch, so the July 12, 2024 entries were made public later than their date. xAI rows use the commit date (UTC). Open-weight rows use the model repository creation month or the model card commit date.

Length: words are counted by splitting the published text on whitespace, so placeholders such as {{currentDateTime}} and template code in Jinja (.j2) files count as words. Counts are approximate and are left blank where the prompt is assembled by code (Gemini CLI) or is template logic (gpt-oss).

Knowledge cutoff: the column records what the prompt says, not an independently confirmed training date. Anthropic's two Claude Haiku 3.5 variants published on October 22, 2024 disagree (July 2024 in the text-only prompt, early December 2024 in the image prompt); both are shown. Grok prompts state no fixed cutoff.

Notable instructions are paraphrased by agitracker.io; quotations are kept under 25 words. Full prompt text is not reproduced. Comparisons between versions were made by diffing the published texts. Vendors can change prompts without updating these pages, so a published version may differ from what a product is running on a given day.

Cite and Reuse This Dataset

The data is free to reuse under CC BY 4.0: copy it, chart it or build on it, with a link back to this page. Download the CSV or JSON.

Citation

agitracker.io (2026). Claude system prompt history and every other officially published AI system prompt. Updated October 29, 2026. https://agitracker.io/data/ai-system-prompts/

HTML link

<a href="https://agitracker.io/data/ai-system-prompts/">Claude system prompt history and every other officially published AI system prompt</a> (agitracker.io)

Frequently Asked Questions

What is the Claude system prompt?

The Claude system prompt is the set of instructions Anthropic places before every conversation in claude.ai and the Claude mobile and desktop apps. It gives Claude the date, its model name and knowledge cutoff, product and tool information, formatting habits and rules for sensitive topics. Anthropic publishes every version on its system prompts page; by October 2026 the prompts ran roughly 3,000 to 4,300 words.

Where can I find Claude's official system prompt?

Anthropic publishes it in its developer documentation under Release notes, System prompts. The overview page links to one page per model, from Claude Haiku 3 to Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, and each page shows dated versions. Anthropic says these prompts apply to the Claude apps, not to the API.

Does the Claude API use the same system prompt as claude.ai?

No. Anthropic states that its published system prompt updates do not apply to the Claude API. When you call the API, the model only receives the system prompt you send. That is why a model can behave differently in the Claude app and in your own application, and why copying the app prompt into an API call is rarely a good idea.

Has OpenAI published the ChatGPT system prompt?

No. As of October 29, 2026, OpenAI has not officially published the system prompt used in ChatGPT. It publishes the Model Spec, the Codex CLI prompts on GitHub, and a default system message in its open-weight gpt-oss chat template that identifies the model as ChatGPT with a June 2024 cutoff. Texts sold or shared as the ChatGPT system prompt are leaks or extractions.

Where are Grok's system prompts published?

xAI publishes them in the xai-org/grok-prompts repository on GitHub, which it created on May 15, 2025. The repository holds prompts for the Grok 3, Grok 4 and Grok 4.1 chat assistants, the @grok bot on X, the Explain this post feature and API safety prefixes. Its commit history shows each change; the latest commit is dated November 17, 2025.

Are leaked system prompts accurate?

Sometimes, but they cannot be verified. Extracted prompts come from asking a model to repeat its instructions, which can produce partial, outdated or invented text, and anyone sharing them can edit them. Some come from test or regional versions. For accurate data, use prompts the vendor publishes, and check the date, because vendors update prompts frequently.

What is the difference between a system prompt and a model spec?

A system prompt is the literal text a product sends to the model at runtime, and it can change daily. A model spec, such as OpenAI's Model Spec, or a constitution, such as Anthropic's, describes intended values and behavior and is used to train and evaluate models. The spec sets principles; the system prompt supplies the date, product facts, tools and deployment-specific rules.

How long is a typical system prompt?

It varies widely. Among officially published prompts, Qwen2.5's default is 12 words and Claude Haiku 3's July 2024 prompt is 114 words, while 2026 Claude app prompts run about 3,000 to 4,300 words. Grok 4's chat prompt is about 1,180 words including template code, and Mistral Small 3.2's recommended prompt is about 400 words.

Related Datasets

• AI Model Release Timeline: Every Major LLM From GPT-1 to GPT-6 (2018-2026): Release dates, open-weight status, context windows, input types and launch API prices for 159 notable LLMs from 2018 to October 2026, each sourced to the lab.
• AI hallucination examples: 65 documented cases with real consequences: 65 sourced cases where AI systems produced false information with real consequences, from court sanctions to refunds and retractions, plus benchmark hallucination rates.

Version History

• October 29, 2026: First published with 81 records, each linked to its source.

Spotted an error or a missing record? Every row links to its source; corrections are made and dated on this page. See all datasets.