Skip to content

ChatGPT · Guide · 3 min read · Last verified 17 Sept 2026

When to let ChatGPT do the task for you

ChatGPT's agent mode can carry out multi-step tasks on the web on your behalf, filling forms, comparing options across sites, booking or purchasing when authorised, rather than just answering questions. Use it for well-defined, low-risk, repetitive tasks, and stay hands-on for anything involving payment details, legal commitments or brand-sensitive public actions.

Free PDF

When to let ChatGPT do the task for you, as a printable PDF

Enter your email and the download opens straight away. You also get occasional free workshops, tools and prompts, unsubscribe any time.

What it is good at

  • Handles multi-step web tasks, compare, gather, fill in, without you clicking through each site yourself.
  • Good at repetitive research tasks like pulling the same data point from many product pages.
  • Shows you its steps as it goes, so you can interrupt before it takes an action you disagree with.
  • Saves genuine time on tasks that are tedious but not judgement-heavy.

Where it falls short

  • It can misinterpret a page and take a wrong action, so anything irreversible needs a human check before confirming.
  • Sites with heavy anti-bot protection or unusual layouts can cause it to fail partway through.
  • It is slower per task than doing something simple yourself; the payoff is on multi-step, repetitive tasks, not quick lookups.

How to use it for creative work

  1. 01

    Start with tasks that have no financial or legal consequence

    Use it first for research aggregation or form-filling with dummy data before trusting it with anything that spends money or commits you to something.

  2. 02

    Give it explicit stopping points

    Say 'stop and ask me before submitting or confirming any purchase' so it pauses at the moment of real consequence rather than completing the whole chain unattended.

  3. 03

    Watch the first run in full

    Don't leave the room on the first attempt at a new type of task, watch what it clicks and how it interprets ambiguous instructions so you can correct the prompt for next time.

  4. 04

    Keep the task narrow and specific

    'Compare shipping costs for this item across these five specific sites' works far better than a broad, open-ended goal it has to interpret from scratch.

  5. 05

    Review its summary against the actual result

    After it finishes, check the outcome directly rather than trusting its own report of what it did, especially for anything customer- or client-facing.

Ready-to-use prompts

Bounded comparison task
Go to these three retailer sites: [links]. For each, find the price, shipping cost and delivery time for [product]. Do not add anything to a cart or enter any payment details. Report back as a table.
Form-filling with a stop point
Fill in this form with the details below, but stop before clicking submit and show me the completed form for approval first: [details].
Repetitive data pull
Visit each of these [number] supplier pages: [links]. Extract the minimum order quantity and lead time from each and put it into a table.
Scheduling assistance
Check availability for [service/venue] across these dates: [dates], and list which are available, but do not book anything without confirming with me first.
Guarded purchase task
Find the lowest current price for [specific product] across these sites: [links]. Stop and ask me before adding anything to a basket or entering payment information.

Questions people ask

Is ChatGPT agent mode safe for online purchases?
It can be used for purchases, but it is safer to set an explicit stopping point before payment so you confirm the final action yourself, since it can misread a page and make a wrong selection.
What tasks is ChatGPT agent mode actually good for?
Well-defined, repetitive, multi-step tasks with low consequences if something goes slightly wrong, comparing prices across sites, filling in known forms, gathering the same data point from many pages.
Can ChatGPT agent mode get stuck?
Yes, unusual page layouts, login walls or anti-bot protections can cause it to fail partway through a task, so always check the outcome rather than assuming completion.

Your next step

Go from free tool to real results

Get free workshops, tools and prompts

Occasional updates from Fashion AI School: free live workshops, new AI tools and new prompts. No spam, unsubscribe any time.