How to Write Better AI Prompts for Business, and How to Check the Result

Most weak AI answers have a plain cause: the request left out something the writer already knew. An AI model cannot see your business, your customer or your rules unless the prompt says so. It fills the gaps with guesses, and you may not notice which parts were guesses until a customer does.

This guide covers two skills: writing a prompt that gives the AI what it needs, and checking what comes back before anyone relies on it. It includes four small tests we ran, templates you can copy, and a section on when a person must review the draft. It is about writing prompts, not about choosing a tool. If you are still deciding which assistant to use, start with how to choose an AI tool and the AI tool stack for a one-person business.

How to read the evidence labels in this guide

Vendor guidance changes. Everything marked official was checked on 29 October 2026, and you should check the current pages before you rely on it.

What "better" means for a business prompt

A better prompt is not a longer or cleverer one. For business work it is a prompt whose draft meets the rules you care about with less fixing. Anthropic's guidance on testing prompts says to define specific, measurable success criteria first and then test against them, instead of judging by feel. OFFICIAL SOURCE ONLY.

Our recommendation: before you write the prompt, write four to six yes/no checks. For a reply to a customer enquiry they might look like this.

The four tests later in this guide work this way. Each comparison has a rule list that we wrote before we ran the prompts.

The AI Spark prompt brief: six parts

This is our own planning model. It is not an industry standard and it is not taken from any vendor. It is a checklist of what business prompts most often leave out. Anthropic and OpenAI both advise clear instructions, relevant context, a stated output format and testing, and the brief packages those ideas for everyday business jobs. OFFICIAL SOURCE ONLY for the vendor advice; the six-part packaging is ours.

Figure 1: The six parts of the AI Spark prompt brief (our own model).

Both vendors suggest marking where pasted material starts and ends, with labelled sections or tags, so the AI does not mix up your instructions with the text it is working on. OFFICIAL SOURCE ONLY.

You may notice what is missing: a line such as "act as an expert". Anthropic's guide describes giving the model a role as a way to focus tone, and says one sentence is enough. OFFICIAL SOURCE ONLY. We did not test roles, so in our brief a role is an optional line inside Context. NOT TESTED.

From a vague prompt to a structured one

ILLUSTRATIVE EXAMPLE, NOT RUN. The task is tidying a rough note before it goes into a CRM. We have not run this particular example, so we do not describe what it would produce.

Bad / vague prompt

Clean up this note: [paste the note]

Structured version

GOAL
Rewrite the rough note below as a clean CRM note. A person will read it before it is saved.

CONTEXT
The note was typed quickly after a phone call with a prospective customer of [type of business].

INPUT
Note: [paste the note]

RULES
- Use only what is in the note. Do not add details, dates or amounts.
- If the next step or its date is missing, write "not stated".
- No more than 6 lines.

OUTPUT FORMAT
Four lines starting with: What happened / What they want / Next step / Date for next step.

CHECK
Before you answer, confirm every line is supported by the note.

What changed

• The goal says where the result goes and that a person will read it.
• The AI is told to use only the note, so gaps are reported instead of filled.
• Missing information has a fixed wording ("not stated"), which makes gaps easy to spot.
• The format is fixed, so every note looks the same.
• A check step asks the AI to compare its draft with the note.

Figure 2: The route a business prompt should take. Steps 4 and 5 stay with a person.

What we tested: four single-run prompt tests

TESTED BY AI SPARK. On 29 October 2026 we ran four controlled single-run prompt tests in Claude, inside an agent working session. Each compares a basic prompt (A) with a structured prompt (B) for the same business task, using the same input. The exact model that answered was not independently verified, so we do not name one.

Every input is made up. "Northfield Bookkeeping" is a fictional business, and Priya, Dana and Omar are invented. No client data was used and none of this is a case study.

We wrote the rules for each task before running it, then counted what can be observed: word counts, format, and whether each rule was met. We did not score quality. Prompt A could not meet rules it never saw, so this is not a contest between equals, and it is not a benchmark, a blind test or a statistical result.

Short version: in each run, B met more of our written rules than A. B was not perfect (it missed one instruction in Test 1), and in Test 2 the basic summary was already good enough for a person to read.

The tables list only the checks we wrote before each run. "Not applicable" means prompt A was never told that rule or format.

Test 1: reply to a customer enquiry

Test 2: summarise rough meeting notes

Test 3: draft a social media post

Test 4: classify customer enquiries

What changed, in plain terms

• Test 1: B was shorter and kept to the facts we gave it. A answered "Yes, we offer monthly bookkeeping" without being told what the business offers. We had not written checks for two more things we noticed in A: it said the firm can "usually get new clients started quickly", and it left [Your name] and [Business name] to fill in. B did not fully meet "exactly two questions": it asked both, then offered the call as a third question.
• Test 2: both summaries were accurate and easy to read. B used the headings we asked for and wrote "no date given" for Omar's task, and A did not mention the missing date. If you only read the summary yourself, the basic prompt was enough. If another system or a project log reads it, fixed headings help.
• Test 3: A added emojis, four hashtags, the word "stress" and benefits we never supplied ("more time to focus", "accurate, up-to-date books"). Its call to action offered two actions. B stayed inside the facts we allowed, but it is also plainer, and plain is a matter of taste.
• Test 4: A was given no categories or format, so it invented its own labels in a table. B returned JSON that parsed and used only the four labels. It filed the complaint under "support" and flagged it for a person, which you might judge differently.

Full prompts and unedited outputs, so you can check our reading. Where prompt B repeated the same input as prompt A, the input is shown once in brackets:

Write a reply to this customer enquiry.

Enquiry: Hi, I run a small cafe in Leeds and my bookkeeping is a mess. Do you do monthly bookkeeping? Roughly how much, and how soon could you start? Thanks, Priya

Subject: Re: Monthly bookkeeping enquiry

Hi Priya,

Thanks for getting in touch, and please don't worry. Plenty of cafe owners find the books pile up when the days are spent running the shop.

Yes, we offer monthly bookkeeping for small businesses. The price depends on things like how many transactions you have each month and which software you use, so I'd like to learn a little more about your cafe before giving you a figure. We can usually get new clients started quickly once we've agreed what's needed, and I'd be happy to give you a firmer start date after a short chat.

Would you be free for a quick call this week? In the meantime, it would help to know:
- Roughly how many transactions you process each month
- Whether you use any accounting software (for example Xero or QuickBooks)
- How far behind the books are

Kind regards,
[Your name]
[Business name]

GOAL
Draft a reply to the enquiry below. A person will review it before sending.

CONTEXT
We are Northfield Bookkeeping (a fictional business used for this test). We provide monthly bookkeeping for small businesses in the UK. We do not quote prices by message. We give a quote after a 15-minute call, and we cannot promise a start date before that call.

INPUT
Enquiry: [same enquiry as in Prompt A]

RULES
- Under 110 words.
- Do not mention a price, a price range or a start date.
- Do not invent services, qualifications or experience.
- Plain English, warm, no exclamation marks.
- Ask exactly two questions: roughly how many transactions per month, and which accounting software (if any) they use.
- Offer a 15-minute call.
- Sign off as Sam.

OUTPUT FORMAT
A subject line on the first line, then the email body. Nothing else.

CHECK
Before you answer, check the draft against every rule and fix any that fail.

Subject: Your bookkeeping enquiry

Hi Priya,

Thanks for getting in touch. Yes, we provide monthly bookkeeping for small businesses, and we would be glad to help sort out your cafe's books.

We don't quote by message, but we can give you a clear answer on cost and timing after a 15-minute call. To prepare for it, could you tell me:

1. Roughly how many transactions go through the business each month?
2. Which accounting software, if any, do you use?

Would a short call this week suit you?

Kind regards,
Sam
Northfield Bookkeeping

Summarise these meeting notes.

Notes: call w/ Dana re website relaunch. launch was 12th now maybe 19th bc photos late. Dana to send photos by fri. Omar to fix contact form bug - says its 'quick'. budget - Dana worried, 2nd invoice not yet approved?? dont know. need decision on blog section - keep or drop, Dana undecided. next call tues 10am

**Meeting summary: website relaunch call with Dana**

**Key points**
- The relaunch was originally planned for the 12th but may now move to the 19th because the photos are late.
- Dana is worried about the budget, and it is unclear whether the second invoice has been approved.
- A decision is still needed on whether to keep or drop the blog section. Dana is undecided.

**Action items**
- Dana: send the photos by Friday.
- Omar: fix the contact form bug (he described it as quick).

**Next steps**
- Next call: Tuesday at 10am.

GOAL
Turn rough notes into a short summary that can be pasted into a project log.

CONTEXT
These are notes from a call about a website relaunch. Dana is the client and Omar is the developer. The notes are rough and some facts are uncertain.

INPUT
Notes: [same notes as in Prompt A]

RULES
- Use only information in the notes.
- Do not guess dates, owners or amounts.
- If something is uncertain or missing, put it under "Unclear in the notes", not under Decisions or Actions.
- Under 120 words.

OUTPUT FORMAT
Four headings in this order: Decisions, Actions, Open questions, Unclear in the notes.
Write each action as: owner - task - due date, or "no date given".
Write "None" under a heading with nothing to report.

CHECK
Before you answer, confirm every action has an owner and either a due date or "no date given".

**Decisions**
- Next call: Tuesday at 10am.

**Actions**
- Dana - send the photos - Friday
- Omar - fix the contact form bug - no date given

**Open questions**
- Keep or drop the blog section? Dana is undecided.

**Unclear in the notes**
- Launch date: it was the 12th and is "maybe" the 19th because the photos are late.
- Whether the second invoice has been approved.
- Dana's budget concern: no detail given.

Write a LinkedIn post announcing our new monthly bookkeeping package for small businesses.

📊 Big news for small business owners!

We're excited to announce our new Monthly Bookkeeping Package, designed to take the stress out of managing your finances.

Running a business is hard enough. You shouldn't have to spend your evenings buried in receipts and spreadsheets. With our package, you get:

✅ Accurate, up-to-date books every month
✅ More time to focus on what you do best
✅ Clear reports so you always know where you stand

Ready to get your finances in order? Send us a message or drop a comment below to learn more.

#SmallBusiness #Bookkeeping #Entrepreneurs #Accounting

GOAL
Draft a LinkedIn post announcing a new monthly bookkeeping package.

CONTEXT
We are Northfield Bookkeeping (a fictional business used for this test). Our audience is owners of small shops, cafes and trades who keep their own books.
Facts we may state: it is a monthly service, and the first step is a free 15-minute call.
Facts we must NOT state: prices, time saved, money saved, results, or client numbers.

RULES
- Under 80 words.
- No emojis.
- At most 2 hashtags.
- Exactly one call to action: ask readers to message us to book the call.
- Plain English.
- Do not use the words "stress", "effortless" or "game-changer".

OUTPUT FORMAT
The post text only.

CHECK
Before you answer, check the post against every rule.

If your bookkeeping tends to wait until you have a spare evening, this may be for you.

We have launched a monthly bookkeeping package for owners of small shops, cafes and trades. It is a regular monthly service, and the first step is a free 15-minute call to see whether it fits your business.

Message us to book the call.

#SmallBusiness #Bookkeeping

Classify these customer enquiries into categories.

1. Hi, do you offer payroll as well as bookkeeping?
2. I was charged twice for last month's invoice.
3. The login link in your email does not open for me.
4. Can you recommend a good accountant for my limited company tax return?
5. Not happy with how this was handled. Please call me.
6. We would like to book a call about starting monthly bookkeeping in January.

| # | Enquiry | Category |
|---|---|---|
| 1 | Do you offer payroll as well as bookkeeping? | Service enquiry |
| 2 | Charged twice for last month's invoice | Billing issue |
| 3 | Login link in your email does not open | Technical support |
| 4 | Can you recommend a good accountant for my limited company tax return? | Referral request |
| 5 | Not happy with how this was handled, please call me | Complaint |
| 6 | Would like to book a call about starting monthly bookkeeping in January | Sales lead |

GOAL
Sort each customer enquiry into one category so it can be routed.

CONTEXT
We are Northfield Bookkeeping (a fictional business used for this test). Allowed categories:
- sales: a question about our services or a wish to start working with us
- billing: invoices, payments, refunds
- support: problems using something we sent or provide
- other: anything else

INPUT
[the same six enquiries as in Prompt A]

RULES
- Use only the four allowed category names, in lower case.
- Set needs_human to true for any complaint, any unclear message, or any message that asks for a call-back about a problem. Otherwise false.
- Do not invent facts about the sender.
- Keep each reason to 12 words or fewer.

OUTPUT FORMAT
Valid JSON only, no other text: an array with one object per enquiry, with the keys id (number), category, needs_human (true or false) and reason.

CHECK
Before you answer, confirm there is one object for each of the 6 enquiries and that every category is one of the four allowed names.

[
  {"id": 1, "category": "sales", "needs_human": false, "reason": "Asks whether payroll is offered alongside bookkeeping."},
  {"id": 2, "category": "billing", "needs_human": false, "reason": "Reports a duplicate charge on last month's invoice."},
  {"id": 3, "category": "support", "needs_human": false, "reason": "An emailed login link does not open."},
  {"id": 4, "category": "other", "needs_human": false, "reason": "Asks for an accountant recommendation, not one of our services."},
  {"id": 5, "category": "support", "needs_human": true, "reason": "Complaint about how something was handled; wants a call."},
  {"id": 6, "category": "sales", "needs_human": false, "reason": "Wants a call about starting monthly bookkeeping in January."}
]

What these tests do not show

• One run per prompt. Outputs can vary between runs and models. Run the prompts again in your own tool before you rely on them.
• One assistant. We used Claude only. We did not test ChatGPT, Gemini or any other tool.
• Not independent. The same working session wrote the prompts, ran them and checked the results. A fresh chat might differ, so repeat the tests with your own inputs.
• Made-up tasks, not client work. We cannot say how much time, if any, structured prompts save, so this guide makes no such claim.
• Our rules, our choices. Different rules would give different results.

Prompt templates for common business jobs

Each template uses [square brackets] for the parts you fill in. Replace them with your own facts. Do not paste passwords, API keys or customers' private details unless your tool's privacy settings and your own policy allow it.

EXAMPLE TEMPLATE. Templates 1 to 3 follow the structure of the prompts we tested. Templates 4 to 6 are NOT TESTED. Try each one on your own examples before you rely on it.

1. Lead reply draft

GOAL
Draft a reply to the enquiry below for a person to review. Do not send it.

CONTEXT
We are [business name]. We offer [services]. [What we do not say by message, for example: we quote only after a call.]

INPUT
Enquiry: [paste the enquiry]

RULES
- Under [number] words.
- Do not state prices, dates or promises that are not in this prompt.
- Ask only these questions: [question 1], [question 2].
- Plain English. Sign off as [name].

OUTPUT FORMAT
A subject line, then the email body. Nothing else.

CHECK
Check every rule before you answer and fix any that fail.

2. Meeting summary

GOAL
Turn the rough notes below into a short summary for our project log.

CONTEXT
Call about [project]. [Person] is the client and [person] is the developer. Some facts in the notes are uncertain.

INPUT
Notes: [paste the notes]

RULES
- Use only information in the notes. Do not guess dates, owners or amounts.
- Put uncertain or missing items under "Unclear in the notes".
- Under [number] words.

OUTPUT FORMAT
Headings in this order: Decisions, Actions, Open questions, Unclear in the notes.
Write each action as: owner - task - due date, or "no date given".

CHECK
Confirm every action has an owner and a due date or "no date given".

3. Enquiry classification

GOAL
Sort each enquiry below into one category so it can be routed.

CONTEXT
Allowed categories: [category 1: what it covers], [category 2: what it covers], [category 3: what it covers], other.

INPUT
Enquiries, numbered: [paste the enquiries]

RULES
- Use only the allowed category names.
- Set needs_human to true for any complaint, any unclear message, or any request for a call-back about a problem.
- Reasons: [number] words or fewer.

OUTPUT FORMAT
Valid JSON only: an array with one object per enquiry, with the keys id, category, needs_human and reason.

CHECK
Confirm there is one object per enquiry and every category is allowed.

4. CRM note cleanup (not tested)

Use the structured version from the section above.

5. Support reply draft (not tested)

GOAL
Draft a reply to the customer message below for a person to approve. Do not send it.

CONTEXT
Our policy on this topic is below. Use only this policy.
Policy: [paste the policy text]

INPUT
Customer message: [paste the message]

RULES
- Use only the policy text. If the policy does not cover the question, write "NEEDS HUMAN: [what is missing]" instead of guessing.
- Do not offer refunds, compensation, legal statements or exceptions.
- Calm, plain English, under [number] words.

OUTPUT FORMAT
The draft reply, or the NEEDS HUMAN line. Nothing else.

CHECK
Confirm every statement in the draft comes from the policy text.

6. Automation troubleshooting (not tested)

GOAL
Help me work out why an automation step failed.

INPUT
What the step should do: [description]
The error message: [paste it, with any keys, passwords or customer details removed]
What changed recently: [description]

RULES
- Give at most three likely causes, most likely first, each with one check I can do.
- Say what you cannot tell from the information I gave.
- Do not tell me to delete data or change credentials without telling me to back up or ask the owner first.

OUTPUT FORMAT
A numbered list.

CHECK
Confirm each cause is linked to something in my description.

How to check a prompt before you rely on it

OUR RECOMMENDATION, built on ideas from the official guides listed under Sources.

1. Write the checks first. Four to six yes/no items, as in the table earlier.
2. Collect a few inputs, including an awkward one. Anthropic's guide on building evaluations suggests test cases that mirror real inputs and include edge cases such as empty, off-topic or hostile messages. OFFICIAL SOURCE ONLY.
3. Run each input more than once. OpenAI notes that outputs vary between runs, and Anthropic describes inconsistent answers across repeated runs as a warning sign. OFFICIAL SOURCE ONLY.
4. Have a person mark each draft against the checks. Count what failed and fix the prompt, not just the draft.
5. Save the prompt with its date, tool and model name. Both vendors note that models differ and change, and OpenAI suggests versioning prompts and pinning model versions in production. OFFICIAL SOURCE ONLY.
6. Re-test when the tool or model changes. Anthropic says model-specific techniques were measured on particular models and should be re-checked on yours. OFFICIAL SOURCE ONLY.

Where a prompt sits inside an automation

CONCEPTUAL / NOT TESTED END TO END BY AI SPARK. Once a prompt works by hand, it can be placed in one step of an automation so that a draft is ready when a lead arrives. Below is a design sketch. We have not built or run an AI-powered lead reply workflow, so this is not a result.

Figure 3: A concept for AI-drafted lead replies. Not tested end to end by Ai Spark.

What we have actually run. Our own n8n test, execution #2689, had no AI step. A form submission was saved to Google Sheets first, then a Gmail draft was created, and no email was sent. TESTED BY AI SPARK, without AI. We mention it for one reason: it shows why a draft stage is useful, because a person can read the draft before anything reaches the customer. The full walk-through is in our n8n lead follow-up guide.

Our recommendation: ask the AI for a fixed format and have the next step check that the format matches before it uses the result. OpenAI's safety guidance also suggests limiting the range of inputs and outputs, and returning vetted content instead of new text where that is possible. OFFICIAL SOURCE ONLY. If you are deciding which tasks are worth automating at all, the automation audit guide covers that choice.

When not to trust a prompt without review

Most prompt mistakes are small and easy to spot. The ones to guard against are those where a wrong sentence costs a customer or you something. Vendor documentation does not claim these tools can be made error-free. Anthropic says its techniques reduce made-up answers but do not eliminate them, and advises validating critical information. OpenAI recommends human review before outputs are used, especially in high-stakes cases, and notes that prompting narrows risk without removing it. OFFICIAL SOURCE ONLY.

Give the AI a safe way out. Anthropic's guidance says that permission to say it does not know can reduce invented answers, which is why template 5 tells the AI to write "NEEDS HUMAN" instead of guessing. OFFICIAL SOURCE ONLY.

What we have and have not tested

Frequently asked questions

A sensible next step

Pick one repetitive task. Write five yes/no checks for a good result, write the prompt using the six-part brief, and run it on three examples of your own. Keep a person on the last step. If you would like to talk through a task you are thinking of automating, you can contact us.

Sources

• Claude Platform Docs: Prompt engineering overview (checked 29 Oct 2026)
• Claude Platform Docs: Prompting best practices (checked 29 Oct 2026)
• Claude Platform Docs: Define success criteria and build evaluations (checked 29 Oct 2026)
• Claude Platform Docs: Reduce hallucinations (checked 29 Oct 2026)
• Claude Platform Docs: Be clear and direct (checked 29 Oct 2026)
• OpenAI API Docs: Prompt engineering (checked 29 Oct 2026)
• OpenAI API Docs: Safety best practices (checked 29 Oct 2026)

Ai Spark is not affiliated with Anthropic or OpenAI. Last reviewed 29 October 2026. Our own tests are described in the section "What we tested: four single-run prompt tests".