Most people who try AI for real work type one sentence, get something vague back, and decide it isn't ready. The tool usually isn't the problem. The handoff is.
If you want to know how to use Claude and ChatGPT for business tasks that actually matter, stop asking and start delegating. Write the prompt the way you'd brief a sharp new hire on their first day.
I call it the delegation prompt. Below is the structure, the choice between chat and a browser agent, and four complete prompts you can copy for common ops work.
Why one-line prompts fail
"Clean up my customer list" sounds clear to you. You know which list, which columns matter, what a duplicate looks like at your company, and what you plan to do with it after.
The AI knows none of that. So it guesses. Guesses look confident, which is worse than looking wrong.
A good delegation prompt removes the guessing. It takes a couple extra minutes to write, and you can reuse it every week.
The delegation prompt: six parts
| Part | What it answers | Example line |
|---|---|---|
| Goal | What outcome do I want, and why? | "I need a clean list so we can send a spring tune-up offer." |
| Context | What does it need to know about us? | "We're a 9-person HVAC company. Customers are homeowners in St. Louis County." |
| Access | What can it look at or use? | "Use only the attached CSV. Don't browse." |
| Constraints | What must it not do? | "Don't delete rows. Flag them instead." |
| Definition of done | What does finished look like? | "A CSV with these columns, plus a summary of changes." |
| Checkpoints | When does it stop and ask me? | "Before you merge anything, show me the first 10 matches." |
Checkpoints are the part most people skip. They're also what keeps an agent from finishing the wrong task very confidently.
Chat or an agent in the browser?
There are now two ways to hand work to these tools, and picking the right one matters.
- Use chat when you can upload or paste everything the AI needs, and you want text, a file or a plan back. Spreadsheets, drafts, summaries and analysis all fit here.
- Use an agent in the browser when the work lives inside a website or app you'd otherwise have to click through. Checking prices on supplier portals, updating records in a CRM, pulling numbers from a dashboard.
On the browser side, Anthropic's >Claude in Chrome can read, click and fill in forms on sites you're signed into, on paid plans as of this writing. OpenAI's >ChatGPT Work mode can use a >cloud browser that asks you to confirm before actions like payments.
Either way, the rule is the same. You sign in yourself. You never paste a password into a chat. The agent stops and asks before it sends, publishes, pays or deletes.
How to use Claude and ChatGPT for business tasks: four prompts to copy
Each of these uses the six parts. Swap the [bracketed] placeholders for your details. The business examples are illustrative.
1. Vendor price check (agent in the browser)
Say you run an HVAC company and want to know whether your usual supplier is still the cheapest on the filters and parts you buy most.
Prompt: Compare vendor prices
Goal: Find out which supplier has the best current price on our top items so I can decide where to order next month.
Context: We are a [9]-person HVAC company in St. Louis. We buy from [Supplier A], [Supplier B] and [Supplier C]. I am already signed into all three in this browser.
Access: Use only those three supplier sites. Do not visit other sites.
Items: [Item 1 with part number], [Item 2 with part number], [Item 3 with part number].
Constraints: Do not add anything to a cart. Do not place orders. Do not change any account settings. If a site asks you to log in or solve a CAPTCHA, stop and tell me.
Definition of done: A table with item, supplier, unit price, pack size, price per unit, and in-stock yes or no. Add the page link for each price. Note anything that looked like a contract or member price.
Checkpoint: After the first item, show me the row so I can confirm you found the right product before you do the rest.2. Clean up a CRM export (chat)
Every CRM collects duplicates, typos and dead contacts. Export it, upload the file, and let the AI do the first pass.
Prompt: Clean a CRM export
Goal: Clean this customer export so we can import it back and send a seasonal offer.
Context: Attached is a CSV export from [CRM name]. We are a [commercial cleaning] company. A duplicate is the same company or the same email, even if spelled differently.
Access: Use only the attached file.
Constraints: Do not delete any row. Do not invent missing data. Do not change email addresses except to fix obvious formatting like extra spaces.
Definition of done:
1. A cleaned CSV with the same columns as the original plus a new "status" column (keep, merge, review).
2. A separate list of suspected duplicates with the rows you would merge and why.
3. A short summary: how many rows, how many duplicates, how many missing emails or phones.
Checkpoint: Before building the final file, show me 10 suspected duplicate pairs so I can confirm your rules match how we think about duplicates.3. Draft a follow-up sequence (chat)
Leads go cold because nobody follows up twice. Have the AI draft the sequence. You approve every word before anything gets loaded into your email tool.
Prompt: Draft a quote follow-up sequence
Goal: Write a short email follow-up sequence for people who got a quote from us and haven't replied.
Context: We are [Company], a [roofing] company in [St. Louis]. Typical job is [describe]. Our tone is plain, friendly and local. No hype. Here are two emails I've written that sound like us: [paste two past emails].
Constraints: Three emails max. Under 120 words each. No fake urgency, no discounts unless I add one. Sign as [Name]. Use [phone number] as the only phone number.
Definition of done: Three emails with subject lines and the day each should send (counting from the day the quote went out). One line on the purpose of each email.
Checkpoint: This is a draft only. Do not send anything or connect to any email tool. I will review and edit before anything goes out.4. Build a weekly report (agent or chat)
If someone pulls numbers from three tools every Friday, that's a perfect handoff.
Prompt: Build the Friday numbers report
Goal: Give me a one-page weekly snapshot every Friday so I can see how the week went in two minutes.
Context: We are a [property management] company. The numbers live in [accounting tool], [CRM] and [scheduling tool]. I am signed into all three in this browser.
Access: Read only. Look at reports and dashboards. Do not edit, export customer data to any outside site, or change settings.
Pull these for the last 7 days and the 7 days before: [new leads], [jobs completed], [invoices sent and total], [invoices over 30 days past due], [open work orders].
Definition of done: A short table with this week, last week and the change. Under it, three bullet points on what stands out. Show where each number came from.
Checkpoint: If any number looks off by more than [50] percent from last week, stop and tell me before writing the summary. It might be a report filter, not a real change.What to do when the work comes back
Treat the output like a new hire's first week. Check a sample against the source. Look at the items it flagged as unsure. Fix the prompt, not just the answer, so next week goes better.
When a prompt works three weeks in a row, it's basically a standard operating procedure. Turn it into one with the SOP template for agents. And if you're still deciding which tasks to hand off first, start with whatever your team retypes most, using the copy-paste tax worksheet.
Frequently asked questions
What is a delegation prompt?
It's a set of instructions written the way you'd brief a new hire: the goal, the background, what tools they can use, the rules, what finished looks like, and when to stop and check with you. It's longer than a normal prompt, and the extra two minutes of writing saves an hour of back and forth.
Should I use Claude or ChatGPT?
Both are capable for everyday business work, and both now offer ways to work inside a browser on paid plans. Pick the one your team will actually use. The delegation prompt works the same in either. I use both, often on the same task, and compare.
Can I give the AI my logins so it can finish the task?
No. Never paste a password, API key or card number into a chat. Sign in to the tool yourself, let the agent work inside that signed-in browser, and approve anything that sends, publishes, pays or deletes. Give it a user seat with limited permissions when you can.
How do I know the output is right?
Build the check into the prompt. Ask for a short list of anything it was unsure about, sources for every number, and a summary of what changed. Then spot-check a few items yourself. If it's going to a customer or touching money, a person reviews all of it.
Or hand me the whole list
Writing good delegation prompts is a skill, and reviewing the output every week is a job. If you'd rather not do either, my agents work your task list, I write the instructions and check every piece, and we meet weekly. Start with a $99 first month.
