Getting started with ChatGPT: your first useful workflow

I would make the first session deliberately small. One account. One page. One person who can tell whether the output is useful. A sprawling strategy exercise gives a beginner too many ways to confuse an attractive answer with a correct one.

My starting rule is the reviewable first run: choose a job whose mistakes you can spot. You are learning how to work with the product as much as asking it to do the work.

Choose a task you can judge

The first task should have an answer you can evaluate without a research project of your own. Summarize a meeting you attended, reorganize a short document or create a brief from notes you have already read. Your familiarity with the inputs is the quality control. Beginning with an unfamiliar market forecast hands the same system both the exam and the marking scheme.

I would use a meeting brief for fictional Cedar Metrics. The inputs are a short account note, a call transcript excerpt and a dated product description. The output is one page for an account executive preparing for a conversation. Success means the rep can separate known facts from unanswered questions without rereading every source.

Keep the first run independent of live CRM writes or outbound messages. That is a workflow choice, not a permanent limitation. You are isolating the part you want to learn: can the product turn these inputs into a useful, faithful result?

Check the workspace before adding context

Sign in through the official product and confirm whether you are in a personal account or an organization workspace. That choice affects available tools, policies and who administers access. If you cannot find a feature described in a guide, inspect the account and workspace before assuming you missed a hidden button.

For company work, use the environment approved by your organization. A source can be technically uploadable and still be inappropriate for a personal workspace. Ask the owner of the data when the rule is unclear. For a first experiment, a fictional account packet removes this dependency and makes the exercise easier to share internally.

Then check that the files you attach are the intended versions. A file called final-notes-new-final is familiar office archaeology; it is poor source control. Rename the packet with the account, date and purpose so the next reviewer can understand what was supplied.

Choose Chat or Work for this run

Chat is useful while you are clarifying what belongs in the brief. Ask which questions the source material can answer and what is missing. Work is appropriate when you want the agent to carry the task through to a finished document or other reviewable file.

You can ask for a small answer in Chat and remain there if it meets the need. A new user does not have to exercise every mode to complete a simple task. Move to Work when the deliverable, multiple sources or tool use makes delegation useful. The goal is fewer handoffs between the request and the accepted output.

If you are on desktop, confirm whether the run uses cloud or local work. Uploaded files and connected sources differ from a folder on your computer. A local file path typed into a cloud chat does not automatically make that file available. Attach it or choose the appropriate environment.

Write a brief with a finish line

The useful prompt contains five things: the outcome, audience, sources, constraints and checks. These are ordinary instructions you would give a colleague. They beat a long paragraph asking the model to be brilliant, meticulous and world-class, because they tell it what those adjectives would mean here.

For Cedar Metrics, specify that the reader is an account executive, the meeting is tomorrow and the brief should fit on one page. Ask for the customer's stated goal, supported facts, unresolved questions and suggested next steps. Tell it to mark missing information rather than complete the story from general knowledge.

Finally, define the output format. A readable answer in the conversation may be enough; a downloadable document may be easier to share. If you want a file, say so. “Make this useful” leaves a surprisingly large gap between your picture of finished and the agent's picture of finished.

Use an example that exposes guessing

Here is a fictional source packet: the account has 120 employees, wants to reduce reporting work and has not confirmed a budget. A transcript says the operations lead will invite finance to a later meeting. Nothing says finance has approved a purchase. The distinction is small enough to disappear in a confident summary and large enough to distort the next call.

Ask ChatGPT to produce separate headings for facts, unknowns and questions. A good brief says that budget is unconfirmed and finance involvement is a next step. A weak one promotes “finance will attend” into “finance is supportive.” You now have a specific error to look for, rather than a vague impression that the writing sounds good.

I would keep this kind of counterexample in the first prompt. It teaches the required standard without prescribing the entire answer. The model still has room to organize the brief, while the business meaning of missing evidence remains explicit.

Inspect the output, not the progress summary

When the task finishes, open the document or read the full answer. A message saying the brief is ready does not establish that it has the right account name, correct dates or the requested uncertainty labels. The file is the deliverable; the completion message is a receipt.

Trace a few important claims back to the supplied sources. Check the highest-consequence details first: buying stage, budget, stakeholder role and next meeting. Then inspect the formatting. Can the rep find the unanswered questions quickly? Are the sources legible? Does the page still work when opened in the application where the team will use it?

Keep factual and editorial feedback separate. “The budget is unknown” corrects meaning. “Move the questions above the history” corrects structure. Mixing both into “make it better” forces another round of guessing. Specific feedback makes the next run useful faster.

Revise one failure at a time

If the output overstates stakeholder support, point to that sentence and the source that contradicts it. Ask for the specific correction and require the rest of the brief to be checked for the same error. That is a better revision than starting a fresh conversation with a longer list of adjectives.

If the brief is too long, preserve the source references and uncertainty labels while shortening background. If a source was unavailable, add it or accept a narrower scope. The important habit is diagnosis: missing context, weak reasoning, wrong format and unavailable tools are different problems.

After a successful revision, ask for the reusable brief instructions. Save them with the source requirements and acceptance checks. You are building a small process that another operator can repeat, not collecting the transcript of one unusually lucky afternoon.

Where the first workflow usually breaks

The broadest failure is trying to solve too much at once. Account research, persona selection, message writing and CRM updates create four review problems. A beginner gets more signal from one faithful brief than from a giant automation whose weak step is difficult to locate.

Another failure is using a source packet that contradicts itself without saying which document is current. Give the newer date or ask the agent to flag conflicts. The model should not silently decide that the more polished document outranks the more recent one. Conflicting positioning is a business decision to resolve, not a formatting nuisance.

The last failure is never checking the second run. A prompt that works on an account with complete notes may guess on an account with missing fields. Repeat the same process with a sparse packet before sharing it. Then use Projects to keep the durable context together. What would you need to see before trusting this brief for a real meeting?

How to set it up

Frequently asked questions

Sources & further reading

• Official quickstart
• Using Chat and Work
• Prompting guidance
• Working with files

ChatGPT and Codex change quickly. This page was last reviewed September 22, 2026; verify time-sensitive details against the official docs above before relying on them.