OpenAI has published a new prompting guide that treats Chat, ChatGPT Work and Codex within a single framework. Its core is not a technique for writing long instructions. It is a way of thinking: first state the outcome you want, then add the context, output format and boundaries the task needs, scaled to the size of the job. ChatGPT, once a conversational partner, now consults multiple sources, creates files and changes code. As its role expands, prompts are starting to shift from a tool for specifying how to answer into a tool for agreeing on what to complete and where to stop.
The four elements are not a mandatory template
For large or important tasks, the new guide lists four elements: Goal, Context, Output and Boundaries. Goal is what you want done. Context is the information or sources that change the result. Output is the format, length and level of detail. Boundaries covers what must not change and which actions require confirmation before they are carried out. You do not have to fill in all four every time, however. OpenAI says a short request is often enough and states plainly that you should use only the elements that help.
This is also not a simple rewording of the beginner-focused material that OpenAI Academy published on April 10, 2026. That material set out three steps: make the task clear, supply useful context and specify the ideal output. It advised splitting work with multiple parts into smaller pieces, while also cautioning against piling on too much detail. The new guide keeps those basics and adds boundary-setting, final checks and mid-task course correction on the same page.
The April material also encouraged users to adjust their request after seeing the first response. The new guide carries that iterative approach forward and makes boundary-setting, mid-task redirection and pre-completion checks more concrete. The flow is clearer: start short, then add missing information as you refine, rather than trying to write a finished set of long instructions up front.
Fewer steps, but boundaries and verification stay
The guide's strongest message is to write the outcome before a detailed process. Specify steps only when the process itself matters. Otherwise, let ChatGPT search, compare information and adjust how it proceeds. If the intended audience or the finished format affects the result, state them explicitly. The requester shows what should be completed and lets the AI choose how to investigate, rather than fixing the search procedure in advance.
Context is also chosen for its effect, not its volume. OpenAI asks users to attach only materials that could change the answer and to say what should be drawn from each. When using connected sources such as Google Drive or Slack, it is enough to indicate where to look and why. There is no need to list search terms or the order in which to read things. Throwing in large amounts of material is no substitute for a carefully written request.
Boundaries that could lead to accidents, on the other hand, should be spelled out. Official examples include "Do not change the approved dates or budget," "Use only the sources provided, and flag gaps rather than guessing," and "Keep messages as drafts and do not send them." OpenAI recommends narrowing boundaries to the one or two that matter most. The idea is to prioritize instructions that prevent accidental changes to dates or budgets and that stop work just before sending or publishing.
Checking is part of the definition of done. For post-meeting action items, have the model verify that every item has an owner and a deadline. For research, have it confirm that no unverified information remains. A model's self-check is not a substitute for human review, though. OpenAI itself asks people to review important outputs before using or sharing them.
One request design linking Chat, Work and Codex
Product reorganization lies behind this unified guide. On July 9, 2026, OpenAI announced ChatGPT Work, which investigates and analyzes across multiple apps and files and produces documents, spreadsheets and presentations. The same day, it described a new desktop app that combines Chat, Work and Codex. Chat answers questions, Work carries out long tasks and Codex handles code, so their uses differ. Still, they share the need to convey what to complete, what to consult and before which actions to stop.
The official guide sorts them this way: use Chat for short questions and rewording text, and use Work for jobs involving multiple sources or tools, multi-step processes or large deliverables. In Work, you convey the desired outcome, the materials, the audience and how to verify the result. If the method affects the outcome, ask for a plan, and have Work wait for approval before sending, publishing or changing information that others depend on.
The basics are the same in Codex, but what gets verified becomes more concrete. You give the desired behavior, the relevant code or reproduction steps, the constraints to respect and how to confirm the change. If you realize mid-run that information is missing, Steer lets you add instructions to the current task, while Queue lets you hold them for the next one. In Codex CLI, Enter is assigned to Steer and Tab to Queue. Rather than inflating the first request, the approach is to watch the work and supply information at the right moment.
A short request is not a sloppy one
"Write the result first" does not mean the specification can be vague. In work where the process determines whether the result is legitimate, such as legal compliance, audits, reproducible experiments and data migration, steps and evidence trails are part of the completion criteria. The new guide also says explicitly to ask for a plan when the method matters. What can be shortened is the means the AI is free to choose, not the constraints the requester is responsible for.
For the same reason, a prompt with all four elements filled in is not a quality guarantee. If the input materials are wrong, the result will fall apart, and even with a final-check instruction, oversights can happen. What helps in practice is not a formulaic long text but a request that has criteria for judging the result, values that must not change, approval before publication and a way to verify the output.
OpenAI's new guidance does not make prompt engineering unnecessary; it changes what users decide. There will be fewer occasions to script the model's entire process and more to decide what it should produce and which actions a person must approve. Going forward, what will set tools apart is less whether a short request yields an attractive first draft than whether Work and Codex can keep to constraints through a long task, reliably stop before approval and return verifiable results.