When should you use it?
- You understand a 1040 well enough to judge the output
- You want a second, independent path to the same number
- You value a documented trail behind every deduction
Not for: Filing directly from the AI; the author used tax software to submit; People who cannot tell a wrong checkbox from a right one
What do you need, and what do you get?
You need
- Wage statements, 1099 forms and the annual brokerage statement
- Bank, card and order-history exports
- Empty federal forms plus the official instruction booklets
You get
- Plain-text working files where each figure traces to a source
- A formula-driven workbook of the whole return
- An audit file pairing expenses with statement lines and written rationale
- A reconciled result between two preparation paths
How does it work, step by step?
- Feed sources, not memoryPut forms, statements and official instructions in folders and have the assistant extract figures to files you can open.
- Let it ask about ambiguityExpect follow-up questions on unclear rows instead of guesses, and answer them in writing.
- Build a traceable workbookGenerate a spreadsheet of the return with formulas so every number links back to its input.
- Prepare it a second wayRun the same return through your usual tax software.
- Reconcile until they matchInvestigate every difference; each one either exposes an AI miss or something the software skipped.
The prompt
Our own prompt, inspired by Geoffrey Doempke and written to work in Claude. Values in lime are placeholders.
Act as my tax preparation assistant for [tax year]. Read every file in /tax/[year]: forms, 1099s, statements and the official instructions. Extract each figure to plain-text working files with the source file and line noted. Ask me about anything ambiguous instead of guessing. Then build a spreadsheet of the return with formulas linking each line to its inputs, and keep an audit file pairing each deduction with its evidence and a one-line rationale. Do not file anything. Flag every checkbox and election for my review.
Inspired by Geoffrey Doempke: I Did My Taxes with Claude. Here's What Happened.. The prompt above is written by playbooks.bot in our own words; nothing from the original is quoted.
What can go wrong?
- When it filled forms directly it ticked the wrong filing status even though the math was right; check every box, not only totals.
- Extra time to submit the paperwork does not mean extra time to pay; send your estimated payment by the April deadline.
Our take
Geoffrey Doempke, who once worked as an individual tax professional at PwC, dropped tax forms, 1099s, bank exports and order histories into folders for Claude to classify and extract into plain-text files. It built a formula-linked workbook and caught a missed credit, then he ran TurboTax too and reconciled both versions until they matched. This is among the most responsible AI tax write-ups we have read, because he designed in a second opinion rather than trusting either path. The reconciliation step is where learning happened, and the audit file built along the way may outlast the return itself. It assumes real tax literacy, so most people should treat the workflow as preparation for their own review or their preparer's, not a replacement. Remember the checkbox mistake: correct math can still sit under a wrong election. Feed sources, not memory. Let it ask about ambiguity. Keep the signed return with a human preparer if you are unsure.
Where the time goes back: The extraction, categorizing and paper trail are the tedious parts; the reconciliation turns two imperfect preparers into one result you can sign with confidence.
FAQ
Which step did he trust the AI with end to end?
Why did he submit through tax software instead?
Source and credit
Inspired by Geoffrey Doempke. We describe the idea in our own words, link to the original, and wrote the prompt ourselves; nothing is quoted.
- Inspired by
- I Did My Taxes with Claude. Here's What Happened.
- Creator
- Geoffrey Doempke
- Published on
- doempke.com
- Original date
- Apr 15, 2026
- License
- Not openly licensed (the poster keeps all rights), so we quote nothing and link to the original
- Originally for
- Claude
- AI dry run
- Oct 23, 2026: passed. Dry-run tested by our AI test agent on sample input. PASS: Gates held on desk check; nothing was sent, filed, purchased or published.
AI dry run on sample data; not run with live connected accounts. The playbooks.bot test agent filled the prompt with synthetic inputs, produced the output and checked it against the page's stated outputs and guardrails. - Verified
- Oct 23, 2026: source link resolves to the original, the credited creator published it, our excerpt or prompt matches it, and the license allows our use.
Are you the creator? Claim or correct this page.