How to Use ChatGPT Effectively at Work: A Practical Guide

I rebuilt our YouTube performance dashboard from a screenshot I found in Google Images, with live figures pulled out of the channel and out of our data warehouse, and it took 17 minutes. The old route was a ticket to the BI team and a two-to-three week wait, because they were fielding requests from every department.

None of that came from a cleverer prompt. It came from handing over a whole job rather than asking a question, and that's the move that decides how to use ChatGPT effectively at work. A prompt gets you a response you then have to act on. A brief gets you a finished thing.

The rest of this guide is the method: how to brief ChatGPT the way you'd brief a colleague, when to use the chat window and when to use ChatGPT Work, which jobs are worth handing over first, how to check what comes back, and how to save a job so you only ever brief it once.

From asking questions to delegating jobs

A prompt gets you a response, and a brief gets you a deliverable. That distinction stayed academic for as long as the tool could only hand you text in a chat window, and ChatGPT Work is what changed it.

The gap shows up in the data. McKinsey's 2025 State of AI survey found 88% of organisations now use AI in at least one business function, while only 39% report any EBIT impact at enterprise level. The tools arrived, but the change in how the work gets done mostly didn't.

ChatGPT Work doesn't sit and wait for your next instruction. You give it a goal, it plans the steps, works through them, checks its own output and has another go when it gets something wrong. That behaviour has a name: software that runs a whole job to completion is an agent. Its sessions run far longer than a chat, because the end product is a finished file rather than an answer you still have to act on.

So the test to run over your week is whether you can describe the deliverable. If you can describe it, you can delegate it. "What should go in a cohort analysis" is a question for the chat window, whereas "build me a cohort analysis from these five exports, with every cell as a traceable formula" is a brief, and it gives you back a workbook.

One practical consequence: use the desktop app. ChatGPT Work runs in the browser too, but the browser version can't read files on your machine, and file access is where most of the value sits. On a set of five CSV exports running to hundreds of thousands of rows, the browser version refused the upload outright, while the desktop app read the same files off my disk without uploading them anywhere.

The four parts of a brief: goal, context, output and boundaries

OpenAI's own guidance splits a good brief into four parts, and all four go in before you press enter rather than getting drip-fed afterwards.

Goal. What result do you need? Not the topic, the outcome.

Context. Which files does it need, and what does it need to know about your business, your customers and your naming conventions? Which tools should it read from and write to? You can wire those up with a plugin, which is an add-on that lets ChatGPT work inside another app, or with an MCP connection, an open standard for pointing an AI tool at a system you already run.

Output. What should the finished thing look like? A workbook, a document, a dashboard, a folder of images.

Boundaries. This is the part people skip, and it's the part that heads off the expensive mistakes. You'd give a new starter guardrails on their first day, so give the agent the same. The template I use: go and delete the duplicates in the CRM, but if both duplicates have an associated deal, flag it and delete nothing.

Briefing a new employee is a better mental model than querying a search engine, and a new employee asks clarifying questions when the brief is thin. So does a well-briefed agent. In one of our own workflows, a branded-asset job stopped before doing anything and asked whether the event was a free session or a paid workshop, because the answer changed the design and the brief hadn't specified it.

Boundaries are also where you protect quality. On that cohort analysis, the rule was that every number had to be traceable: no hardcoded values, formulas only, so the logic could be audited cell by cell. That one constraint is the difference between a spreadsheet you can defend in a board meeting and one you have to take on trust.

When to use the chat window and when to use ChatGPT Work

Both have a job, and picking the wrong one is what makes the tool look either slow or shallow.

The chat window is a thinking partner. Brainstorming, drafting, quick analysis, talking through a decision, explaining something you don't understand yet. You want an answer in seconds and you're going to act on it yourself.

ChatGPT Work is a colleague you hand a task to, much the same way Claude separates its chat window from Claude Cowork. It reads your files, uses your connected tools and produces the deliverable, and it takes as long as it takes. On that cohort analysis across five systems, it ran for over an hour before returning the workbook.

There's a second choice inside ChatGPT Work: run the job on your own computer, or run it in the cloud. Anything touching local files has to run on your machine, which also means the machine has to stay awake. Jobs that only use connected apps can run on OpenAI's servers instead, and those keep going after you close the laptop.

Worth knowing before you budget for this: ChatGPT Work is available on all plans on desktop, including the free tier, with web and mobile access reserved for Plus, Pro, Business, Enterprise and Edu. Free and Go get limited access, so you'll hit usage caps quickly, but you can test whether any of this holds up before you pay for anything.

Five jobs worth handing over this week

The pattern that works is picking something you already redo often, rather than inventing a new project. These are all real 9x workflows.

Reporting you keep rebuilding. That YouTube dashboard came together in 17 minutes because ChatGPT Work could pull live figures from the channel and from BigQuery through connectors into our data tools, which are the ready-made links that let an AI tool read from an app you already run.

Analysis that's slow by hand. A cohort analysis I ran pulled five exports from four systems into one workbook, with retention split by acquisition channel and every figure left as an editable formula. Claude does the same class of work through its own file creation environment, so this isn't specific to one vendor.

Bulk edits across a tool with no API. Adding a tracked call to action to more than 130 YouTube descriptions, each needing its own dynamic short link, with different rules depending on whether the description already opened with a lead magnet, no CTA, or a subscribe line. That's hours of clicking, which is why it kept sliding down the list.

Research runs. A lead-research job produced 25 companies showing AI-training buying signals, with the source URL for every signal so each one could be checked. It ran for over 20 minutes and split itself up by region.

Small admin that quietly eats your week. A receipts folder where photographing an expense and dropping it in gets it renamed, filed by date and added as a row in a spreadsheet.

What these have in common is that each one has a describable deliverable, each was already being done by hand, and each has a clear finish line.

What to do once the job comes back

Verify the output before you use it

Delegating doesn't mean accepting. Every job that comes back needs a check proportionate to what you plan to do with it, and that check is a system rather than a personality trait. In the same McKinsey research, the organisations getting the most value from AI were more likely to have defined processes for when model outputs need human validation.

Build the check into the brief where you can. Traceable formulas mean you can audit the maths, and a source URL against every research claim means you can spot-check the ones you're going to repeat out loud. On the 25-company lead run, every signal carried the link it came from, so a claim about a company opening a principal AI transformation role could be confirmed in one click.

Expect to iterate. That dashboard came back with a subscriber graph missing both axis labels and a broken column in the videos table. I said so, and five minutes later both were fixed. A first version being imperfect is normal, and it's cheap to correct.

The rule of thumb is that the further the output travels, the harder you check. Something you'll read yourself needs a skim, while something going to a client, a board or a newsletter list gets read line by line.

Save the job so you only brief it once

Once a job comes back the way you wanted, you can save the instructions so the agent repeats it without being talked through it again. That saved set of instructions is a skill, and running one means supplying the inputs rather than the method. This is where using ChatGPT effectively stops being a personal habit and starts compounding.

Do the task first and save it afterwards, because the obvious route round is the wrong one. ChatGPT offers a skill creator that invites you to describe the skill you want from a blank page, and doing it that way risks spending an hour writing instructions for something the tool can't actually do. Run through the task manually with the agent instead, verify and refine until you're happy with the output, and only then ask it to save what you just did.

Two things follow from that, and they're worth knowing early.

Skills make jobs cheaper as well as faster. When part of a task is deterministic and needs no reasoning, the skill writes itself a small script to handle that part, so the model runs the script instead of working the answer out again on every pass. That cuts token use, meaning the volume of text the model has to read and write, which is what you're billed on. Anthropic's documentation on how skills load describes the same mechanism on the Claude side: a short description of the skill stays in front of the model, the full instructions load only when the skill is triggered, and bundled scripts run without their code ever taking up room in the model's working memory, which is the context it can see at any one moment.

Skills can call other skills. An agent can hand parts of a job to copies of itself, which is how sub-agents split a job in parallel, and skills compose in much the same way. One of our skills cuts a single clip from a video given the timestamps. A second reads a transcript, works out where the eight segments start and stop, and calls the first one eight times. Build the smallest reliable piece first and then compose, and on the third run of anything, ask whether this should be two skills instead of one.

The team version is where the return sits. A domain expert writes the skill once and the rest of the team runs it, and new starters inherit a folder of skills that already encode how your company does things, which shortens the ramp considerably.

Six things to avoid when you hand over a job

Treating it as a faster chat. If you're using ChatGPT Work to ask questions, you've picked the slow tool for the job. Use the chat window.

Under-briefing, then blaming the output. A thin brief produces a generic result, and the four parts take two minutes to write.

Assuming every feature is on your plan. Scheduled tasks, for instance, are not included on the free plan, so check before you design a workflow around one.

Not saving the workflow. If you brief the same job twice, you've wasted the second brief. Save it as a skill and it runs the same way each time, which also cuts token use, because the deterministic parts get written into a small script rather than reasoned out afresh on every run.

Letting the tool invent your format when you already have one. If your team has a proposal structure or a reporting layout that works, load it as a template, so the agent fills your model rather than inventing its own.

Starting with the hardest thing you do. Start with the most repeated thing instead. Momentum comes from a job finishing.

Start using ChatGPT effectively today

Pick the task you redo most often this week and write the four-part brief for it: goal, context, output, boundaries. Run it on the desktop app so it can reach your files. Check what comes back against a rule you set in the brief, correct it once, then save it as a skill, so the second run costs you a set of inputs and nothing else.

It's a small change to make: stop using ChatGPT as the place where you ask questions, and start using it as the thing that does the work. Getting a whole team past the question-asking stage is the part that takes structure, and our training programmes do that by function, using your own tools and your own workflows.