You ask an AI tool to write a customer message, it gives you something with the wrong name and an invented delivery date, and you fix it by hand. Next week the same thing happens. If that sounds familiar, the problem isn't the tool alone. You aren't recording what goes wrong, so you keep paying for the same mistake with your time and data.
A mistake log fixes that. It's a simple sheet where you note each failure, guess the cause, and write the prompt change that worked. After a few weeks you'll have a short list of rules that make your results better. This guide shows you how to set one up in about ten minutes.
Why a log works better than "trying again"
When an answer is bad, most people retype the request with a few extra words and hope. That sometimes works, but you learn nothing you can reuse. A log forces you to name the failure, and named failures repeat in patterns.
Typical patterns are: invented facts, wrong tone, ignored length limits, wrong format, mixing up details you pasted in, and weak local-language output. Once you see which pattern is yours, you can add a standing line to your prompts that prevents it.
It also protects you when work is paid. If a client finds an error, you can show how you check outputs. That is a real selling point for freelancers and small-business helpers.
Set up the log in Google Sheets
Use a free spreadsheet. It works on a phone, saves automatically, and you can open it offline if you enable offline access beforehand. Create one sheet with these columns:
- Date
- Task (for example, "reply to unpaid invoice", "summarise meeting notes")
- What I asked (paste the prompt, or a short version)
- What went wrong (one sentence)
- Type of error (pick from a fixed list, see below)
- Likely cause
- Fix that worked
- Cost (minutes lost to fixing it)
Add a drop-down for "Type of error" using data validation, so you don't invent new labels every time. Keeping labels consistent is what lets you count them later.
A starter list of error types
| Error type | What it looks like | Usual fix |
|---|---|---|
| Invented fact | A date, price, law or source that you never supplied | Tell it to use only the text you provide and to say "not stated" otherwise |
| Wrong tone | Too stiff, too pushy, or too casual for a customer | Name the reader and give a short sample of your own writing |
| Ignored limit | Too long, or more points than you asked for | State the limit first, then repeat it at the end |
| Wrong format | Paragraphs when you wanted a table | Show the exact column headings you want |
| Mixed-up details | Swaps names or amounts between customers | Process one customer at a time |
| Weak local language | Stiff or wrong Yoruba, Igbo, Hausa, Pidgin or Swahili | Give a glossary and have a native speaker review |
| Missing context | Generic answer that ignores your situation | Add who you are, who the reader is, and the goal |
If you work with local languages, our client glossary guide shows how to build the reference list that fixes the last-but-one row.
How to log a mistake in under a minute
Do it right when you spot the error, not at the end of the week. By then you'll have forgotten what you asked.
- Copy your prompt into the sheet. Remove customer names, phone numbers, account details and anything private first. Our post on using AI at work without leaking customer data covers this habit.
- Write what went wrong in plain words: "gave a due date I never mentioned."
- Pick the error type.
- Guess the cause. Was your request vague? Did you paste too much? Did the task need facts the tool couldn't have?
- Try one change only, then record whether it worked. Changing three things at once tells you nothing.
Don't log every small wobble. Log anything that cost you more than a couple of minutes, or anything that could have embarrassed you if it had gone out unchecked.
A worked example (hypothetical)
Say Amaka runs a small fabric shop in Aba and uses an AI chat tool to draft WhatsApp reminders for customers with unpaid balances. Over two weeks she logs eight mistakes:
- Three are "invented fact": the tool added a payment deadline she never gave.
- Two are "wrong tone": messages sounded like a bank threatening legal action.
- Two are "mixed-up details": she pasted five customers at once and amounts got swapped.
- One is "ignored limit": a reminder ran to six lines instead of two.
She also notes that each one cost about four minutes to fix, so roughly 32 minutes in total. Now she has three clear rules to add to her standard prompt:
- "Use only the facts below. If a date isn't given, don't mention one."
- "Tone: friendly and respectful, like a shopkeeper talking to a regular customer. No threats."
- "Write one message for one customer. Maximum two lines."
The next week she logs only one mistake. The numbers are made up, but the method is real: count, find the biggest category, fix that first. For the message-drafting side, see drafting polite follow-ups for unpaid balances.
Turn the log into a reusable prompt checklist
Every two weeks, sort the sheet by error type and count each one. The top two or three are your priorities. For each, write one standing line and keep it in a note on your phone, ready to paste at the top of any new request.
Your personal checklist might look like this:
- Who is the reader, and what do I want them to do?
- Use only the facts I give; say "not stated" if something is missing.
- Length limit and format.
- One item at a time.
- Tone example.
Then review the log again a month later. Delete rules that no longer earn their place, because long prompts have their own problems and cost you typing time and data.
Use the log to judge tools, not just prompts
Some mistakes are not your prompt's fault. If the same error survives your best fixes, that's a limit of the tool for that task. Record it as such. When you're deciding whether to pay for a subscription or switch tools, the log gives you evidence: how often it failed, and how many minutes it cost. Pair it with the method in how to test an AI tool on your own work before paying.
Also mark tasks where the tool should never work alone: legal wording, tax figures, medical advice, visa rules, and anything with someone's money attached. For those, the fix is always a human check against an official source.
Common mistakes with the log itself
- Making it too detailed. Eight columns is plenty. If it takes five minutes per entry, you'll stop.
- Logging private data. Strip names, BVN, NIN, account numbers and addresses before pasting.
- Blaming the tool every time. Check whether your request was vague first.
- Never reviewing it. A log nobody reads is just a diary of frustration.
- Trusting a fix forever. Tools change, so a rule that worked in the past may stop mattering or stop working. Re-test occasionally.
Frequently asked questions
Do I need a spreadsheet, or can I use a notebook?
A notebook works if you only use AI occasionally. A spreadsheet is better once you have 15 or more entries, because you can sort and count by error type.
How many entries before it becomes useful?
About 10 to 15. Patterns usually show up by then, especially if you use AI daily.
Should I log good results too?
Yes, but briefly. Save prompts that worked well in a second tab named "Good prompts" so you can reuse them.
Can this help in a job interview?
It can. Employers hiring for AI-assisted roles like to hear how you check outputs. Showing a log with a few real examples, with private details removed, demonstrates that habit better than saying "I always double-check."
Your next step
Open a new Google Sheet today, add the eight columns, and log the next AI answer that makes you groan. Review it in two weeks and write your first three standing prompt rules.
Results from any side hustle or online work vary widely, and nothing here is a promise of income. Be wary of anyone who asks you to pay upfront to start earning.