You need a complex report. How do you ask ChatGPT to prepare it?

You need to prepare a report to support a decision in your organisation. There are documents to review, information to cross-check and conclusions to justify. You want to use ChatGPT, but where do you start?

Do you type “write me a report”, attach some files and wait? Do you choose a particular model? Do you use Chat or Work? Can you do it from your phone?

Before assigning the work, it helps to know what is available to you. You do not need to learn programming or master every technical term. You need to understand a few decisions that can make a substantial difference to the result.

We will work through them using a single example: a services company wants to choose a document management application. It has three proposals, a budget and a list of requirements. It needs a report that compares the options, identifies what still needs checking and supports a decision.

The goal is not to obtain ten pages that sound professional.

It is to obtain a useful basis for making a decision.

1. First: which account do you have, and what does it allow?

Access to ChatGPT does not mean access to every model, tool and usage allowance.

Current documentation includes Work and Codex in Free, Go, Plus, Pro, Business, Edu and Enterprise plans. It is therefore incorrect to say that you always need to pay to get started. What varies is how much you can do, with which models and under which conditions.

Free and Go let you begin with limited access. Plus and Pro offer greater usage capacity, with significant differences between them. Business adds organisational and administrative features; Enterprise and Edu also depend on workspace terms and settings. Work and Codex share usage allowances. Plans and limits.

Before continuing, check three things:

• The name of your plan.
• The models and modes actually shown in your account.
• Any usage limits or warnings displayed by the application.

In a business account, administrators may restrict certain features. Nor should you assume that the options will be identical across applications.

Do not upgrade simply because a plan looks “more professional”. First identify the specific limitation you need to overcome.

2. From your phone or your computer?

Your phone may be enough to talk to the AI, give instructions and review responses.

A computer is usually more convenient for organising several documents, comparing tables, examining sources and revising a lengthy report. This is a practical difference, not a measure of the intelligence available.

Another distinction matters more: where you give the instruction is not necessarily where the work is executed.

In Work, when cloud execution is available, a task can continue after you close the application or switch off your computer. Local work, by contrast, can use the computer’s files and applications within the access permissions enabled. How Work operates.

Remote access from a phone to a connected computer is also possible when the feature is available and configured. In that case, the computer must remain awake, online and accessible. The phone is the interface used to direct it; it is not performing all the work itself. Remote access.

For our first test, I would start in the browser or desktop application. We do not need to connect the entire organisation to compare three proposals.

3. Chat, Work and Codex are not three names for the same thing

The practical distinction is this:

Chat is for conversation: asking questions, clarifying, exploring an idea or developing a draft through successive exchanges.

Work is designed around delegating an outcome: a report, presentation, analysis or file requiring several steps. It can use authorised materials and tools and produce deliverables you can review.

Codex is particularly oriented towards developing and running software. It can do other work too, but you do not need to start there to prepare your first report. Introduction to Work and its relationship with Codex.

In our example, you could use Chat to understand the proposals and build the report through conversation.

In Work, you could delegate the comparison and request a reviewable document and table as the result.

Later, if you perform this comparison every week, Codex could help you build a tool that organises proposals or checks particular calculations. That would be another step.

The distinction is not “Chat for amateurs and Work for professionals”. The question is which way of working suits the assignment.

4. The model and its thinking intensity are different decisions

The model is the system that processes information. Thinking intensity indicates how much reasoning effort you allow it to devote, within the available options.

OpenAI’s current catalogue includes, among others:

The interface may show simplified power settings or allow you to select the model and effort separately. Levels can include Low, Medium, High, Extra High and Max, depending on the model and environment. Ultra needs a separate explanation: it can incorporate automatic delegation, not merely more reasoning by a single agent. Greater effort can improve the result, but also requires more time and resources. Models and thinking controls.

How does this apply to our report?

To extract each proposal’s price and duration, I would begin with a moderate setting.

To analyse contradictions, dependencies, hidden costs or the risks behind a recommendation, I would try a more capable model or greater effort.

I would not change both at once if I wanted to find out which one improves the result.

And I would distinguish between two instructions:

“Deliver a brief conclusion.”

“Examine the alternatives carefully before recommending one.”

The first describes what I want to receive. The second describes the work I expect behind it. A five-line conclusion may require considerable analysis.

5. Before attaching files, prepare the materials and the brief

For learning, use public, synthetic or explicitly authorised documents for that environment.

A professional account does not replace an assessment of what information you may share. Nor should you interpret “working locally” as a guarantee that the model operates offline or that no data leaves the computer.

Our test will use three fictional proposals and a requirements sheet.

The requirements should explain at least:

• How many people will use the application.
• Which functions are essential.
• The available budget.
• Which constraints must not be breached.
• Who will read the report and what decision they must make.

We can then give this assignment:

I need a report to select a document management application for a services company with twenty people. Use the three proposals and the attached requirements sheet. Do not consult the internet for this first version. Compare functions, implementation, costs over three years, limitations and dependence on the vendor. Distinguish explicitly stated amounts from your own calculations and show your assumptions. Do not invent terms that do not appear in the documents. Identify missing information. Deliver a comparison table and a report of no more than 1,500 words, with a reasoned recommendation and the questions we should resolve before deciding. Identify the document and section supporting each important fact. If none of the proposals provides sufficient grounds to recommend an option, say so. Do not contact vendors or take any external action.

You do not need a novel’s worth of instructions. You need to remove the important ambiguities.

If you do not know how to frame the assignment, you can begin by asking for help:

Before preparing the report, ask me the questions needed to clarify the objective, comparison criteria and outcome I need.

You can also ask for an outline before drafting. Correcting the direction at the start is easier than rewriting twenty pages afterwards.

If we subsequently need external research, we will add it as a second task, defining what should be checked and which sources to use.

One important clarification: writing “do not take external actions” is an instruction, not a technical block. When we connect applications or grant permission to act, we must also configure the corresponding controls.

6. Where do agents come in?

An agent can carry out a task through several steps and tools. A system with subagents can distribute parts of the work and then bring the results together.

For example, in our comparison:

• One agent examines costs and commercial terms.
• Another checks compliance with the requirements.
• A third looks for contradictions and challenges the recommendation.

The coordinator integrates their contributions.

Work offers subagents to accounts eligible for that feature. Availability should not be inferred solely from the plan’s name. At many levels, delegation must be explicitly requested; Ultra can use it proactively. Multi-agent work consumes more resources than a comparable single-agent task. How subagents work.

We should also avoid confusing two things: asking a response to adopt three perspectives does not prove that three independent agents have worked on it. Look at what the system actually executed.

You may not need delegation for three simple proposals. For an extensive analysis with separable parts, it may be worth testing.

More agents do not guarantee greater rigour. They can also multiply mistaken interpretations or duplicate work.

7. Receiving a report does not mean the work is finished

The presentation may be excellent while the recommendation is poorly grounded.

Before using the result, check:

The facts. Do prices, deadlines and terms match the originals?

The calculations. Does the report include implementation, maintenance, additional users and relevant assumptions?

The sources. Does the cited section actually say what the report claims?

The gaps. Does it acknowledge what it does not know, or fill the gaps with assumptions?

The recommendation. Does it address your requirements, or simply choose the proposal that appears most comprehensive?

You can ask the same system for an additional review:

Now review the report for errors, unstated assumptions and conclusions that lack sufficient support. Identify what information could change the recommendation.

This helps, but it is not independent validation. The same system may overlook its own error.

The decision still needs a person who understands what is being decided.

Practical tip: use another AI as an adversarial auditor

In my daily work, when I receive an important report, I first read it and write down my comments. I then commission an adversarial review from another powerful model, such as Anthropic’s Opus.

“Adversarial” does not mean disagreeing on principle. It means actively looking for where the report could fail: errors, omissions, hidden assumptions, overly categorical conclusions or sources that do not support the claims.

I do not ask it to write another report. I ask it to critically examine the one I already have, which is probably a good piece of work.

One possible brief would be:

Review this report as an adversarial auditor. I am also attaching the original assignment, authorised materials and my comments. Look for material errors, contradictions, omissions and insufficiently justified conclusions. Check calculations and, where you have access, open the original sources to verify that they support the claims. If you cannot consult a source, say so: do not treat it as verified. Distinguish demonstrable errors from reasonable objections and mere stylistic preferences. For each issue, identify the affected passage, explain its significance and propose a concrete correction. Do not find faults just for the sake of it, and do not rewrite the whole report. Acknowledge sections that are well supported.

I then read the audit and decide which observations to accept. With that selection, I return to the system that prepared the report—in this example, ChatGPT—and request the changes:

Incorporate my comments and observations 1, 3 and 5 from the audit. Preserve the rest of the report, apart from adjustments needed for consistency. If a correction contradicts the original documents, flag it before applying it. Summarise what you changed at the end.

This is not about starting an endless game of ping-pong between models. It is about introducing a second critical perspective into a process that I continue to direct.

For an ordinary report, one review round and a final check of the changes may be sufficient. If a material issue remains unresolved, it must be checked against the sources or referred to a person with the necessary expertise.

Using another provider does not guarantee independence or accuracy either: two models can share the same error. Before sending the report to the second system, check that its data may also be shared with that provider.

The second AI contributes objections. It does not replace my judgement or become the arbiter of the outcome.

8. Run a test in your organisation

The best way to understand what you need is not to discuss in the abstract which model is most powerful.

It is to compare results on a specific piece of work.

I would set up this small experiment:

First test: Chat

Use the materials and brief above. Record the model and thinking intensity selected.

Save the initial result before correcting it. Then record the clarifications and human work needed to turn it into something useful.

Second test: Work

Repeat the assignment with the same documents and, if available, the same model and effort.

This lets you observe what the way of working contributes: the steps performed, files delivered, traceability or ease of review.

Third test: another model or greater effort

Repeat the task, changing only one variable.

For example, compare Sol with Astra in the same environment if your account allows both. Or keep the model and increase the effort.

If the available options do not permit an equivalent comparison, record that. Do not attribute to the model a difference that could also be due to the tools or execution mode.

Delegation to agents can be a fourth test, not a prerequisite for getting started.

For every test, use a new conversation or task and keep the same starting materials.

Then compare:

• Which result contains fewer significant errors?
• Which identifies missing information better?
• Which provides a stronger justification for its recommendation?
• How long does it take to deliver something usable?
• How much time do you spend verifying and correcting it?
• What usage or limits can you observe?

If you use a second model as an auditor, apply the same review brief to the different results. Its observations can help you compare them, but do not turn its assessment into an automatic ranking.

A single trial does not establish a universal rule. But it can reveal which configuration deserves further testing in your organisation.

You do not need to master everything to get started

You do not need to learn every model, connect every application or build a team of agents on day one.

You need to choose a task, prepare suitable materials, define the outcome and review what happens.

Then you can increase the complexity, one thing at a time.

You may discover that Chat handles much of your work well enough. Work may save you coordination and file-preparation effort. A more powerful model may make a substantial difference to some assignments and add very little to others.

That is exactly what we want to find out.

The process is simple to describe, although learning to do it well takes practice: assign, read, question, correct and decide.

The question is not which account has the most features or which model tops a ranking.

It is this:

What is the simplest, safest and most proportionate way to obtain the outcome my organisation needs?

Features and documentation consulted on 16 September 2026. Availability may change depending on plan, rollout, application and organisational restrictions. The comparative exercise described is a proposed test, not a performance comparison already conducted.