Get tasks doneBusiness size: Solo

A Fukuoka developer had Claude Code build a bookkeeping pipeline that reconciled 200 transactions and flagged its own doubts

Ken Imoto @kenimo49 · Japan · Claude Code, MCP connection to his accounting platform (platform not named)

Go to the skill

What the agent did

Built a re-runnable pipeline that pulls the month's transactions over MCP, matches bank rows to receipts by amount and date, proposes a category with a one-line reason for each row and flags rows it isn't sure about.

Result, as claimed

200 transactions reconciled and 11 miscategorised, a 94.5% hit rate. Review shrank to 18 flagged rows plus a sample. 7 of the 11 errors were in the flagged pile and 4 were caught on the sample.

This is the result as claimed in the public source linked below. It has not been checked and is not a typical outcome.

Proof

Read the public source

Try it yourself

Have the AI sort a month of transactions and tell you which rows it's unsure about, so you check 20 rows instead of 200.

Difficulty: Beginner

Prompt

Here is last month's business bank export with account numbers removed: [paste CSV]. Here is my category list: [list]. Here are my receipts and invoices with date, vendor and amount: [paste list]. For each row, match a receipt if amount and date agree, suggest one category and give a one-line reason. Add a confidence column: SURE or CHECK. Mark CHECK for anything personal-or-business, any meal or entertainment, any payment covering two things and any row with no receipt. Don't guess tax rules. End with a count of SURE and CHECK rows. Return a draft only. Never send, spend, change settings or write to a live system.

Prompt: starter by Solopreneur Code

Steps

  1. Export one month of bank transactions and remove account numbers.
  2. Paste it with your category list and receipt list.
  3. Check every CHECK row yourself.
  4. Spot-check 10 SURE rows. If any are wrong, add a rule and run again.
  5. Give the reviewed sheet to your accountant. Don't file from it directly.

Watch out: The AI can't know if a purchase was personal or business, or your local tax rules, so a confident wrong category can slip into a tax filing.

Starter version by Solopreneur Code

Skill: Money and bookkeeping

Add this skill

Money and bookkeepingDifficulty: Beginner

Bookkeeping draft with doubt flags

Sorts a month of transactions, matches receipts and marks every row SURE or CHECK so you review the doubtful ones first.

Works with
Claude, ChatGPT, Gemini, Any AI chat Any AI chat can work on a pasted CSV. Pulling transactions directly needs a coding agent with an MCP or API connection to your accounting software, as in the story. Not tested.
Tools in the story
Claude Code
Country
Japan

How to use this: Copy the skill, paste it into Claude, ChatGPT or Gemini, and ask it to save it as a skill. If it cannot, it will still follow the steps.

Show skill

If the Copy button does not work, select this text and copy it.

Sign up for the newsletter

Get the next story from Solopreneur Code in your inbox.

Form not loading? Subscribe on Substack

Back to all stories

More like this

Browse all stories
Get tasks doneBusiness size: SoloIncludes prompt

A newsletter writer's Claude Code agent turned screenshots into 38 scheduled Substack Notes he did not write, with no engagement lift

Dheeraj Sharma @genaiunplugged · Country unknown · Claude Code, SubflowAI Substack connector, macOS scheduler, Substack Notes

Each morning a headless Claude Code run takes the oldest screenshot, classifies it, writes a 45 to 115 word Note in his voice, checks it against copy gates, uploads the image, schedules it for the next free day, reads the schedule back and archives the screenshot. Screenshots with emails, phone numbers or orders are set aside.

Result, as claimed

29 Notes scheduled by 23 Sep and 38 by 12 Oct with none written by him. 12 measured Notes had a median of 3.5 likes against 5 for his other Notes, which he reads as no lift.

Get tasks doneBusiness size: CompanyIncludes prompt

A Malaysian point-of-sale company runs about 170 AI agents its co-founder built, and sales admin per lead fell from about 15 minutes to about 1

Fong Wai Hong · Malaysia · Claude Code, Salesforce

When someone books a demo, an agent researches the shop and briefs the caller. Other agents rank every open lead into a live call list that re-sorts about every 30 seconds, so a promised callback rises to the top. After each call the conversation is transcribed and the lead record is filled in. He also replaced a couple of hundred Zapier automations with his own stack in about four weeks.

Result, as claimed

Fong says the average time to close out a lead is now about 1 minute, down from about 15. He says reps make two times more calls a day and meetings booked per 100 leads went up about 25%, which he says shows up in the bottom line. He says the marketing team used to spend 30 to 40% of its time fixing broken Zaps.

Get tasks doneBusiness size: SoloIncludes prompt

A consultant running two companies hands his tax letters, bank queries and company filings to AI agents and says they saved him at least 2 days in a month

Peter Hartree @peterhartree · Country unknown · Claude Code, Codex, Gmail, WhatsApp, Google Drive

He sends a screenshot, PDF or forwarded message with a short request. The agents have access to his Gmail, WhatsApp and Drive and can use a browser. Examples he lists: tracing an unexplained bank seizure to a tax charge, checking a tax portal and posting a document by registered mail, reviewing and approving year-end company accounts, and flagging terms in a merger engagement letter that contradicted what was agreed by email.

Result, as claimed

He estimates the AIs saved him at least 2 days of work over the past month. He says the error rate seems similar to or lower than his own and that agents rival a human PA for many tasks.