"Look it up," "Make it into a document," "A little more detail."
Repeating these three exchanges back and forth for an hour—this is the "correction loop" that people first encounter when they start using ChatGPT Work.
The cause isn't a lack of model performance. It's because the initial request doesn't include "what, for whom, in what format, and to what extent" the task should be completed.
ChatGPT Work isn't a function for answering questions; it's a function for completing deliverables. Therefore, you shouldn't give it questions—you should give it completion criteria.
There are only 7 things you need to provide in that first request:
・Purpose: Why are you making this?
・Deliverable: What is the final product?
・Materials: Which files or apps should it look at?
・Target Audience: Who will see it?
・Constraints: Word count, deadlines, prohibitions.
・Completion Criteria: What must be met for it to be finished?
・Process: Should it check the plan first or go straight to completion?
This 7-item set is provided as a copy-pasteable template at the end of this article. If you want to see it now, skip to the "Copy-Paste Template for the 7-Item Set" section.
In this article, I will compare "requests that lead to re-dos" and "requests that pass on the first try" for each of the 7 items. Then, I'll provide a full example of a request using all 7 items. Finally, I'll cover settings in ChatGPT Work that eliminate the need for subsequent requests altogether.
Requests that require constant re-dos have 7 things in common
Requests that increase re-work are missing information.
In most cases, one of the following is missing:
・Using vague terms like "make it look good"
・Not specifying who the document is for
・Changing the purpose halfway through
・Adding necessary files after the fact
・Changing the deliverable format halfway through
・Not communicating the completion criteria
・Not specifying items that need confirmation
Don't feel bad if this sounds familiar. It's just that your previous habits with ChatGPT are ingrained.
Traditional ChatGPT was based on a one-question, one-answer premise. You send something short, see the response, and send another short prompt. This back-and-forth was actually the correct way to use it.
However, Work has a different design philosophy. It's a feature that gets stronger the fewer exchanges you have, so using it with a back-and-forth premise causes friction. This is the first stumbling block.
ChatGPT Work is a "Completion Specialist," not a "Consultant"
OpenAI's official help documentation clearly distinguishes the roles of Chat and Work.
Chat is described as a feature suitable for quick interactive support and everyday questions. Work, on the other hand, is explained as an agent designed for longer, multi-step tasks and the creation of finished deliverables.
(Source: OpenAI Help Center "ChatGPT Work and Codex" https://help.openai.com/ja-jp/articles/20001275-chatgpt-work-and-codex
In other words, Work is not a partner for brainstorming direction. It is the specialist in charge of taking it to the finish line once the direction is decided.
ChatGPT Work was announced on July 9, 2026 (US time) and is based on GPT-5.6 (Source: OpenAI Official https://openai.com/ja-JP/chatgpt-work/).
The official page explains it like this: It gathers context across tools, files, and desktop apps. It plans the approach. It executes actions. And it creates high-quality spreadsheets, documents, and slides.
"Gather context, plan, and execute." Work handles these three steps on its own.
Conversely, what we must provide are the prerequisites for those three steps. For what purpose, using what materials, and where does it end? If these are blank, Work starts running with blanks.
So the usage is simple:
Solidify things in Chat while you're still undecided. Once the goal is set, hand it over to Work. Switch by saying, "Please create the finished product from here."
This one step alone will visibly reduce re-dos.
The 7 Items to Provide in the First Request
Here is the core content. Let's look at the 7 items in order.
1. Purpose: What is this job intended to decide?
Purpose serves as the criteria for Work to make its own judgments during the process.
A weak request looks like this:
"Research the competition."
With this, Work doesn't know what to prioritize. It has no basis to judge whether it should list features in detail or dig into pricing.
A request that passes on the first try looks like this:
"I want to understand the competitors' price ranges and billing methods in order to determine the sales price for a new product."
When the purpose is included, priorities are set automatically. If it's for deciding a price, it will prioritize pricing pages and billing terms over feature lists. Work will handle this judgment for you.
The trick to writing the purpose is to state in one sentence "what you will decide after seeing this document."
Instead of "I want information," say "to decide the sales price." Instead of "I want to know the current status," say "to decide the budget allocation for the next term." Specifying the target of the decision stabilizes the selection process.
2. Deliverable: Specify file format, volume, and structure
"Summarize it in a document" is the phrase most likely to trigger revisions.
The moment you say "document," it's unclear whether it's a text file, a slide deck, or a spreadsheet. If this isn't decided, you'll end up saying "Actually, make it a slide" after seeing the output.
You should mainly specify the following:
・File format (Slides, Document, Spreadsheet)
・Volume (Number of slides, characters, or rows)
・Structure (What to put on which page)
・Writing style (Polite, concise, casual for internal use)
・Whether to include tables or graphs
・How to write citations
Example of a strong request:
"Please create a 10-slide deck for a management meeting. Slide 1: Conclusion, Slides 2-4: Current Status, Slides 5-7: Causes, Slides 8-9: Improvements, Slide 10: Next Actions. Keep the text on each slide under 120 characters."
Specifying the structure seems tedious. However, this tedium happens only once. The tedium of not specifying the structure happens every time you ask for a revision.
3. Materials: Which apps and folders to look at
Work's strength is gathering materials across multiple apps. The official page states it supports over 1,400 plugins (Source: OpenAI Official https://openai.com/ja-JP/chatgpt-work/).
Slack, Microsoft Teams, Google Drive, SharePoint. In addition to these, email, calendars, CRM, and project management tools are also targets for cross-referencing.
However, being connected is different from looking at the correct materials.
A weak request looks like this:
"Make a document about recent sales."
Work doesn't know when "recent" starts or which file contains the sales figures. If it finds an old document, the final product will be made with old numbers.
A request that passes on the first try looks like this:
"Please check the monthly sales table in the 'FY2026_Monthly_Sales' folder on Google Drive. Also, look at the posts from the last month in the Slack sales channel. Based on that, analyze the cause of this month's sales decline. Use the monthly sales table as the primary source; if it conflicts with Slack comments, prioritize the sales table."
When specifying materials, include these 5 points:
・Which app to look at
・Which folder or channel to look at
・What timeframe to target
・What to trust as the primary source
・Which to prioritize if information conflicts
The last line is effective. Deciding how to handle conflicts in advance prevents Work from getting stuck or choosing one arbitrarily.
4. Target Audience: The deliverable changes based on who reads it
The meaning of "summarize clearly" changes depending on the reader.
Clarity for a frontline staff member is different from clarity for an executive. The former implies detailed procedures, while the latter implies speed to conclusion.
Therefore, write the audience's attributes.
"The readers of this document are 3 executives who are unfamiliar with frontline operations. Please add a one-line footnote for technical terms upon their first appearance. Summarize the numbers needed for judgment on the first page."
Specifying the audience simultaneously determines the writing style, handling of technical terms, and the order of information. Since one line decides three things, it's a very cost-effective item.
For external documents, add conditions like "Use official company names" or "Do not include unconfirmed figures."
5. Constraints: Write what you don't want it to do first
The constraints section is for prohibitions. If this is blank, Work might do unnecessary things out of a sense of helpfulness.
Effective constraints include:
・Do not fill in numbers based on speculation.
・Clearly mark items that could not be confirmed as 'Unconfirmed'.
・Do not reference confidential files.
・Keep text per slide under 120 characters.
・Limit technical terms to 5 or fewer.
The first and second points are worth including every time.
A weakness of generative AI is filling in unknown parts convincingly. By instructing it to "don't fill, write unconfirmed," the subsequent verification work becomes much lighter.
This is because humans only need to check the parts marked as unconfirmed.
6. Completion Criteria: Provide the definition of 'done' upfront
This is the most effective of the 7 items. It's also the one most often omitted.
What's missing in many requests is the condition for ending. "10 slides" is a content specification, not a definition of completion. Even if there are 10 slides, it's not complete if it doesn't communicate the message.
Write completion criteria like this:
・Completion is defined as a state where a first-time reader can understand it in 5 minutes.
・Completion is defined as a state where all numbers have citations.
・Completion is defined as a state where unconfirmed items are clearly marked as 'Unconfirmed'.
・Completion is defined as a state where it can be distributed at the meeting as is.
When you provide completion criteria, Work begins to inspect its own output. You've given it the question: "Does this state meet the completion criteria?"
Conversely, without completion criteria, completion for Work is simply "when all instructed elements are included." Including elements is different from being a usable document. This gap is the true identity of the correction loop.
If you can't write completion criteria, remember how the deliverable will be used. Who will see it, when, and in what situation? Putting that into words becomes your completion criteria.
7. Process: Have it show a plan first or go straight to completion
Finally, specify the work process.
ChatGPT Work has a Plan Mode. According to the official page, in Plan Mode, ChatGPT gathers context, asks questions, and creates a step-by-step plan. Users can suggest changes or approve the plan to start work (Source: OpenAI Official https://openai.com/ja-JP/chatgpt-work/).
For important deliverables, this is worth using.
If you realize the direction was wrong after all the work is done, it's a total re-do. If you catch it at the planning stage, you only fix a few lines.
Write in the request:
"Before execution, please present only the research steps and the structure of the deliverable. Start the main work only after I approve."
The main points to check are:
・Are the information sources correct?
・Are there any missing files?
・Does the structure of the deliverable match your intent?
Conversely, for routine tasks that are unlikely to fail, you can skip the plan confirmation and go straight to completion. Switch based on the importance of the deliverable.
Official help also states that you can check progress, answer questions, change direction, and approve important actions. It's designed so you can choose whether to leave it entirely to the AI or hold the reins at key points.
Copy-Paste Template for the 7-Item Set
Here is a format where you can fill in the 7 items. Replace the [ ] parts.
▼ Copy from here
【Purpose】 The purpose of this job is to decide [ ].
【Final Deliverable】 Please create [ ] as the final product. The format is [ ] and the volume is [ ].
【Target Audience】 The readers of this deliverable are [ ]. Please add footnotes for technical terms upon their first appearance.
【Information to Use】 Please refer to the [ ] folder in the [ ] app. Use [ ] as the primary source; if information conflicts, prioritize [ ].
【Target Scope】 Target information from [Month/Year] onwards.
【Constraints】 Word count, page count, deadline, and prohibitions are [ ]. Do not fill in numbers based on speculation. Clearly mark items that could not be confirmed as 'Unconfirmed'.
【Completion Criteria】 Completion is defined as meeting the state of [ ].
【Process】 Before execution, please present only the work steps and the structure of the deliverable. After approval, proceed to the final form.
▲ End copy
You don't need to fill in every field every time.
For light tasks, just "Purpose," "Deliverable," and "Completion Criteria" are effective. A realistic operation is to use all 7 for heavy deliverables.
Here is a short version with just those 3 items:
▼ Copy from here
【Purpose】 I want to create materials to decide [ ].
【Deliverable】 Please create [ ]. The format is [ ] and the volume is [ ].
【Completion Criteria】 Completion is defined as meeting the state of [ ]. Do not fill in numbers based on speculation; clearly mark unconfirmed items as 'Unconfirmed'.
▲ End copy
Since it's only 3 lines, you can use this regularly for daily requests. Use the 7-item version for meeting materials or external deliverables.
Example: A Complete Request for Competitor Research in One Go
What does it look like when you put the 7 items into an actual request? Let's compare a weak version and a strong version of the same job.
The weak request is this:
"Research the competition and summarize it in a document."
In this one sentence, six things are undecided: the purpose, the number of competitors, the reader, the comparison points, the format, and the end point. Therefore, at least 6 re-dos are guaranteed.
A request that can be handled in one go looks like this:
▼ Copy from here
Please research 5 major domestic competitors to determine the sales strategy for a new product.
Prioritize official websites, pricing pages, and official announcements from the past year. Compare price, target customers, main features, strengths, and weaknesses.
The final deliverable is a 10-slide deck for a management meeting. Slide 1: Conclusion, Slides 2-6: Comparison of each company, Slides 7-8: Market gaps, Slide 9: Our path to winning, Slide 10: Next actions.
The readers are executives who are unfamiliar with frontline operations. Please add a one-line footnote for technical terms upon their first appearance.
Add citations for numbers and do not fill in unconfirmed information with speculation. Clearly mark items that could not be confirmed as 'Unconfirmed'.
Completion is defined as a state where a first-time reader can understand it in 5 minutes.
Before execution, please present only the research steps and slide structure. Start the main work after I approve.
▲ End copy
It looks long, but you only write it once.
Moreover, this request can be reused next time. Just by replacing "5 major domestic competitors" with another target, the same high-precision research will run.
100 revisions have no reusability. One design has reusability. This difference accumulates as a difference in time.
Eliminating the Second Request Onwards
Once the first request passes using the 7 items, the next step is to eliminate the second request onwards using ChatGPT Work's features.
Turn successful requests into scheduled tasks instead of typing them every time
If you are typing the same request every week, you can move it to scheduled tasks.
According to official help, you can create both one-time and recurring tasks, and you can ask to be notified when changes are confirmed or important updates occur. Monitoring tasks can notify you only when there is something worth reporting and can be stopped when completion conditions are met (Source: OpenAI Help Center "Scheduled Tasks in ChatGPT" https://help.openai.com/ja-jp/articles/10291617-scheduled-tasks-in-chatgpt).
The order of progression is:
・First, succeed manually once.
・Save the successful request text as is.
・Automate it after the output quality has stabilized for 2-3 times.
・Don't notify every time; notify only when there is a change.
・Specify the end condition for when to stop.
If you automate immediately, low-precision deliverables will accumulate automatically. Move to automation after it passes manually. It's safer to follow this order.
Limitations are also stated officially. Tasks cannot be executed more than once per hour. The limit for active tasks varies by plan. The same help guide states a maximum of 3 for Go, 5 for Plus, and 15 for Pro, Business, and Enterprise.
One more often overlooked note: If a task is created within a project that has files, that task cannot access the project's files, as stated in the official help. If you want a recurring task to use material files, check this point first.
Hand over the workspace instead of making it search for materials
Work can handle local files and apps via the desktop app. The official page also introduces a new browser feature with multi-tab support.
Here too, if the hand-off is sloppy, time is lost in searching. You should specify the following:
・Target folder
・Files to use
・Files not to use
・Save destination
・Whether to overwrite the original file or save as new
・How to name the file
Specifically, "overwrite or save as new" is worth including every time. If the original data is overwritten, the cost of re-doing it jumps significantly.
Note that availability varies by plan. At the time of the official announcement, the macOS desktop version was available for all plans. The web and mobile versions are being rolled out sequentially to Plus, Pro, Business, Enterprise, and Edu. Check the official page for current availability.
Choose the model based on the weight of the work
ChatGPT Work is based on GPT-5.6, and three models are available: Sol, Terra, and Luna.
Summarizing OpenAI's explanation, the usage is as follows:
・Sol: For complex work where judgment and finish quality are critical.
・Terra: A balanced model for performance, speed, and cost, for daily tasks.
・Luna: High-speed and low-cost, for processing large volumes of routine tasks.
Using the top-tier model for everything isn't necessarily the right answer. Sending simple formatting to Sol is overpriced.
What's realistic is to use Terra as the standard and upgrade to Sol only for important deliverables where quality is lacking. Drop simple tasks like classification or extraction to Luna.
If you tell it the difficulty at the beginning of the request, such as "This is a high-level analysis comparing multiple documents," the accuracy of judgment will also improve.
Humans only need to check 3 places after completion
When you provide the 7 items and specify completion criteria, the verification work also becomes lighter. You don't need to re-read everything.
Look at these 3 places:
・Numbers: Are there citations? Are there any numbers without citations mixed in?
・Proper Nouns: Are company names, product names, and people's names spelled correctly?
・Parts marked 'Unconfirmed': Research only these yourself.
The reason for putting "don't speculate" and "write unconfirmed" in the constraints was to make this verification possible.
If a human has to double-check everything an AI creates, no time is being saved. Have the AI side declare where you should look. Only when you design this far have you truly delegated.
Official case studies also talk about how the use of time has changed.
An NVIDIA representative said that before events, they spent about 40% of their working time on manual aggregation and analysis. Now, because ChatGPT executes that process twice a week, they can secure time for strategy planning and customer response (Source: OpenAI Official https://openai.com/ja-JP/chatgpt-work/).
On the same page, a Virgin Atlantic representative also stated that competitor analysis, which usually took several weeks, is now completed in a few hours.
What both have in common is that they didn't just ask the AI something; they handed over work in a form that can be executed repeatedly.
You don't need to write long prompts
The secret to mastering ChatGPT Work isn't writing long spells.
It's simply providing the Purpose, Deliverable, Materials, Audience, Constraints, Completion Criteria, and Process together at the beginning.
Instead of correcting a vague request 100 times, provide the completion conditions in the first request. Check the plan before execution, and after completion, have humans look at only three things: numbers, proper nouns, and "unconfirmed" items.
ChatGPT Work isn't a function for repeating one-question, one-answer exchanges. It's a function for providing goals and judgment criteria and leaving it to reach the final form.
Doing just one thing today is enough.
In your next request to Work, add the line: "Completion is defined as meeting the state of [ ]." Of the 7 items, this is the most effective. The presence or absence of this one line alone will change the number of back-and-forth exchanges.
