OpenAI's prompt guide starts with roles, not wording

Start here

Someone opening OpenAI's prompt engineering guide today, expecting one clever phrase, instead meets a sequence. The guide, retrieved 16 September 2026, opens with message roles before it ever discusses wording. A developer message carries the application's instructions and is stated to take priority over a user message; the model's own turns are assistant messages. That ordering matters for a reader building a repeatable prompt rather than asking a one-off question: the roles decide what the model treats as a rule versus an input, before any clever phrasing does.

What the documents say

The guide's sections run, in the order the page presents them: message roles and instruction following, keeping prompts as versioned code rather than throwaway text, formatting instructions with Markdown or XML tags, few-shot learning, adding external or proprietary context, and planning for the context window's limit on how much text a prompt can carry. A companion page, the guide to generating text, restates the same role priority in plainer terms: developer messages are prioritized ahead of user messages, and user messages behind them. It also defines prompt engineering as the process of writing instructions that make a model consistently meet a stated requirement. Both pages are OpenAI's own documentation, describing what the vendor ships today, not an outside measurement of it.

Check this

A reader can test the priority claim directly: send the same request twice, once with an instruction placed only in the developer-level field and once mixed into the user turn, then compare whether the model follows a conflicting follow-up request less often in the first case. The guide also separates advice for reasoning models from its other current models under two headings, so a second check is to open both named sections and see which matches the model actually being called, since advice written for one is not guaranteed to transfer to the other.

What holds and what fails

The structural advice, roles, versioning, marked sections, holds regardless of which model answers a request, because it concerns how the request is packaged rather than one model's quirks; this is an editorial reading, since the guide does not rank its own sections by effect size. What fails is treating any instruction as permanent: the address for this guide moved once during this reporting, from platform.openai.com to developers.openai.com, and its code samples name a current model a later one will replace. A saved prompt tuned to today's model-specific behaviour needs a retest once the vendor ships a new model, part of why the guide recommends versioned code over loose text.

• Open the guide's role section and note which field a saved prompt currently uses for its instructions.
• Add one structured example to a repeated prompt and compare the next output against the previous one.
• Bookmark the page by topic rather than by address, since OpenAI has already moved this one once.

None of this requires special vocabulary. It requires reading the vendor's current page, noting which model it was written for, and copying the ordering, roles and structure first, before polishing any wording.