Hands-on: build an end-to-end workflow

What you're building

Imagine this is a job you'll do every month at Fernway: take the latest sales export and the team's meeting notes, and produce a short, honest update for the senior team: the headline numbers, the one thing that needs attention, and a clear note of anything the data can't tell you. You're going to build that as a proper little workflow, using the Level 2 skills together:

• a Project to hold the source files and the standing instructions, so next month you just drop in new files;
• file analysis to clean and total the messy sales data reliably;
• a written brief produced from the analysis and the notes, in your house style;
• a verification pass so you know exactly which figures you checked.

Set aside about 16 minutes with ChatGPT open. You'll need two Fernway files: the Fernway sales CSV and the Fernway meeting notes. Prefer your own material? A real sales or activity export plus a set of team notes works just as well, but strip anything confidential first (Phase 0 rules), especially on a personal account.

Step 1: Set up the Project

Create a new Project and call it something like "Fernway monthly sales update". Give it standing instructions describing the recurring job, so every chat inside starts oriented. Then upload both files to the Project itself (not to a single chat) so any conversation in the Project can draw on them.

This Project produces a monthly sales update for Fernway's senior leadership team, who are busy and non-technical. Each month I add a sales export (CSV) and the team's meeting notes. Standing rules: use UK English, keep everything concise, calculate figures with code rather than estimating them, and always tell me which rows or figures you couldn't verify. Treat the uploaded files as the source of truth; if they don't say something, say so rather than inventing it.

Step 2: Clean and total the data

Start a fresh chat inside the Project. Because the files live at the Project level, you can go straight to the analysis. Run the audit-first approach from the files-and-data lesson: problems before totals.

Using the sales CSV in this Project, first give me a short data-quality report: blank cells, dates in an odd format, region names that look misspelled, and any row where Units times Unit Price doesn't equal the Total. Calculate, don't estimate. Then, after I confirm, correct the obvious region typos, recalculate the mismatched totals from Units times Unit Price, and give me total revenue by Region and by Product, listing any rows you had to exclude.

Read the data-quality report properly before you say "go". On this file it should flag the region typos ("Sotuh", "Midlnads"), the mixed date formats, the blank cells, and (the important ones) the two rows where the recorded Total doesn't match Units times Unit Price. Confirm how you want each handled, then let it produce the corrected totals.

Step 3: Turn it into a brief

Now bring in the notes and ask for the actual deliverable: a short written update that combines the numbers with the context from the meeting.

Using the corrected totals above and the meeting notes in this Project, draft a short update for the senior team: (1) two-sentence headline on how the month went, (2) revenue by region as a short list, (3) the single thing that most needs their attention, drawing on the meeting notes, and (4) one honest line on what the data can't tell us. Mark any figure I haven't yet verified with "[check]" so I know what to confirm before sending.

Step 4: Verify, then finish

This is the step that separates a workflow you can trust from one that just looks finished. Take the figures marked "[check]", and any that aren't, and confirm the load-bearing ones yourself. Open the CSV and spot-check the two corrected totals (the £650-that-should-be-£600 and the £49-that-should-be-£490). Confirm the regional totals add up. Make sure the "needs attention" point really reflects the notes and isn't a plausible embellishment. Then have ChatGPT produce the clean final version.

I've verified the regional totals and confirmed the two corrected rows. Remove the "[check]" markers now. One change: lead with the region that grew most, and keep the whole thing under 150 words. Give me the final version I can paste into an email.

Your checklist

• [ ] I created a Project and gave it standing instructions for the recurring job.
• [ ] I uploaded both source files to the Project, not to a single chat.
• [ ] I ran an audit before any totals, and read the data-quality report.
• [ ] I decided how each data problem was handled, rather than letting the tool decide silently.
• [ ] The brief combined the cleaned numbers with context from the meeting notes.
• [ ] I verified the two corrected totals and the regional figures against the CSV myself.
• [ ] I ended with a concise update I'd actually send, and I know which figures I checked.

How to tell you did it well

• The workflow is reusable. Next month you could drop in new files and re-run the same three prompts. If you'd have to rebuild everything from scratch, the Project setup didn't do its job.
• The numbers are real, and checked. The totals came from calculated code, the two planted errors were caught and corrected, and you personally confirmed the figures that matter; you can name them.
• The brief is faithful. The "needs attention" point traces back to something actually in the meeting notes, not a confident invention. The "what the data can't tell us" line is honest, not a token gesture.
• You stayed in control. At each hand-off (cleaning, totalling, drafting) you decided, rather than accepting whatever appeared.

If any of those slipped, that's the step to repeat. Doing the whole chain once more is how it becomes a five-minute monthly habit.

Try it now

Common mistakes

• Skipping the Project and doing it all in one chat. It works once, but you've built nothing reusable and the files aren't shared. The Project is what makes this a monthly routine instead of a one-off.
• Totalling before auditing. Run the sums before catching the typos and the two mismatched totals and your headline numbers are wrong from the start, and wrong invisibly.
• Letting the brief outrun the data. A confident "sales are up because of the new campaign" is only true if the notes and numbers say so. Keep the draft tied to the sources, and keep the honest "what we can't tell" line in.
• Trusting the finished-looking update because the workflow felt rigorous (over-trust). A neat Project, a cited-looking analysis and a polished brief all feel trustworthy, and that feeling is exactly what lets an uncorrected total or an invented cause slip through to the senior team. The process doesn't verify the numbers; you do. Spot-check the load-bearing figures against the source every single time, no matter how professional the output looks.

Keeping current

Projects, file analysis and how they interact are all evolving, so the exact menus and limits may differ from this walkthrough. When something doesn't match, check OpenAI's help articles on Projects and file uploads, plus the ChatGPT release notes. Accurate as of 13 July 2026.