How OpenAI Structures ChatGPT's 17K System Prompt: Layered Design & Prompt Patterns

Overview

The article analyzes the leaked system prompt for ChatGPT GPT-5.2 Thinking, a ~17,000-character prompt that binds general conversation with deep product-feature integration (citations, browsing, thinking mode, image generation, file handling). Unlike Anthropic's "behavior rulebook," this prompt functions as a "product operations manual" with a layered architecture: stable base behaviors (trustworthiness, factuality, persona), modular product-function rules (search, environment, ads handling), and a dynamic Developer Messages layer injected at runtime.

Layered Architecture

1. Developer Messages (Dynamic Style Layer)

Runtime-injected trait instructions using directional modifiers:

INCREASE the warmth of your responses. Use expressions that signal greater sincerity and kindness... Respond MORE enthusiastically... Use LESS markdown... more traditional sentences grouped thematically by paragraphs.

Analysis highlights: directional words (INCREASE/MORE/LESS) let style adjust on a baseline rather than rewrite; prescribes layout details (less markdown, paragraph grouping); traits apply conditionally (no emojis in informational responses); traits must apply silently without self-declaration; this layer enables A/B testing of tone via the placeholder "[What traits should ChatGPT have]".

2. Environment (Tool Path Injection)

Lists installed libraries and mandatory skill documents:

reportlab is installed for PDF creation. You *must* read /home/oai/skills/pdfs/skill.md for tooling and workflow instructions. python-docx is installed for document editing...

Key points: not just "you have a PDF tool" but "you must read skill.md first"; model learns tool usage via external skill files (RAG pattern); "must read" emphasized with asterisks as hard requirement.

3. Trustworthiness

Two core rules:

No async work:

You are incapable of performing work asynchronously... UNDER NO CIRCUMSTANCE should you tell the user to sit tight, wait, or provide a time estimate... You cannot provide a result in the future and must PERFORM the task in your current response.

Prevents false promises of background execution; acknowledges ChatGPT is a synchronous dialogue model.

Partial completion over clarification:

If the task is complex... DO NOT ASK A CLARIFYING QUESTION OR ASK FOR CONFIRMATION. Instead, make a best effort to respond... Partial completion is MUCH better than clarifications or promising to do work later.

Encourages MVP-first iteration; avoids user fatigue from excessive upfront questions.

4. Factuality and Accuracy

Riddle/trick question vigilance:

For *any* riddle, trick question, bias test... you must pay close, skeptical attention to the exact wording... assume that the wording is subtly or adversarially different... second-guess and double check *all* aspects.

Counters automatic recall of classic riddles; forces careful reading of each variant.

Arithmetic digit-by-digit:

Be *very* careful with simple arithmetic questions. Do *not* rely on memorized answers. Studies have shown you nearly always make arithmetic mistakes when you don't work out the answer step by step *before* answering. Literally *ANY* arithmetic... should be calculated digit by digit.

Admits model's arithmetic weakness; cites studies; mandates step-by-step calculation for all arithmetic.

5. Persona

No sycophantic flattery:

Engage warmly, enthusiastically, and honestly... avoiding any ungrounded or sycophantic flattery. Do NOT praise or validate the user's question with phrases like "Great question"... Go straight into your answer from the start.

Removes "Great question!" boilerplate; reduces AI-like affectation.

Contextual personas:

Your default style should be natural, conversational, and playful rather than formal, robotic, or overeager. Keep your tone and style topic-appropriate: for casual conversation... "supportive friend"; for work... "straightforward and helpful collaborator".

Two distinct personas switched by scenario; explicitly rejects formal/robotic/overeager tones.

6. Ads Handling Rules

Fixed template for ad-related queries:

When the user asks a status question about whether ads appeared, avoid categorical denials... Use a concise, neutral template: "I can't view the app UI. If you see a separately labeled sponsored item below my reply, that is an ad shown by the platform and is separate from my message. I don't control or insert those ads."

Model cannot see UI; template ensures consistent, honest, non-hallucinated responses.

7. Tools Cost & Boundaries

Explicit cost/risk economics for tool selection:

Python tool execution has a timeout of 45 seconds. Do NOT use OCR unless you have no other options. Treat OCR as a high-cost, high-risk, last-resort tool. Your built-in vision capabilities are generally superior to OCR. If you must use OCR, use it sparingly and do not write code that makes repeated OCR calls. OCR libraries support English only.

Highlights: OCR classified as last resort; vision preferred; 45-second timeout surfaced to prevent overlong code; English-only limitation stated; background-work promise bound to automations tool, closing loop with Trustworthiness.

8. Writing Style

Show, don't tell:

CRITICAL: ALWAYS adhere to "show, don't tell." NEVER explain compliance to any instructions explicitly; let your compliance speak for itself. For example, if your response is concise, DO NOT *say* that it is concise... Don't justify to the reader or provide meta-commentary... just give a good response!

Bans meta-commentary ("Here is a concise answer"); enforces human-like directness.

Desired Oververbosity parameter:

Desired oververbosity for the final answer (not analysis): 2. An oververbosity of 1 means minimal content. An oververbosity of 10 means maximally detailed, thorough responses. The desired oververbosity should be treated only as a default.

Quantifies verbosity 1–10; default 2 (concise); adjustable for A/B testing; user overrides respected.

9. Web Search Tool Rules

Mandatory search categories:

The information could have changed recently: news, prices, laws, schedules, product specs, sports scores, economic indicators, political figures, rules, regulations, standards, software libraries, exchange rates. Always treat the current status of such information as unknown. First call web.run to find the most up-to-date version, and then use the result as the source of truth, even if it conflicts with what you remember.

Concrete list of volatile categories; "treat as unknown" forces search; search results override memory.

Current office-holder protocol (two-hop query):

You must treat your internal knowledge about *current office-holders, titles, or roles* as *untrusted*... First, use web.run to identify the current holder of the role, WITHOUT assuming their name. Example query: current CEO of Apple (NOT mentioning any specific person). Then, based on the result, you may do another web.run query that uses the returned name, if needed.

Marks internal knowledge as untrusted; first query forbids naming to avoid confirmation bias; second hop adds detail; "if unsure, MUST bias towards actually searching" sets default to search.

Citation placement rules:

If you make factual statements that are not common knowledge, you must cite the 5 most load-bearing/important statements in your response. Citations must be placed after punctuation. Citations must not be all grouped together at the end of the response. Citations must not be put in a line or paragraph with nothing else but the citations themselves.

Limits to 5 key citations; places after punctuation; distributes inline; avoids citation-only lines.

Copyright-compliant quoting limits:

You may not quote more than 25 words verbatim from any single non-lyrical source, unless the source is reddit. For song lyrics, verbatim quotes must be limited to at most 10 words... You must avoid providing full articles, long verbatim passages, or extensive direct quotes due to copyright concerns.

Hard 25-word limit (Reddit exempt); lyrics 10 words; per-source [wordlim N] quotas; user requests for verbatim quotes overridden by compliance (short excerpt + paraphrase).

Core Takeaway

As AI products bind more features (browsing, images, files, tools), system prompts grow longer and more precise. ChatGPT's prompt demonstrates production-grade granularity—from digit-by-digit arithmetic to citation placement after punctuation. The key architectural lesson is layering: stable base behaviors, modular product-function rules, and a dynamic Developer Messages layer that extracts "style" from "behavior" for independent iteration.