Build a tax-law research assistant from public documents
Martin Alderson had Claude Code research about 10,000 UK tax documents; he scored its answer to a past exam question 2.5 of 3, where plain Opus failed.
Done with Claude Code
What they did
Martin Alderson had Claude Code write a program that downloaded UK tax law and the tax office's official guidance manuals. That gave him about 10,000 text documents in folders. He then rewrote Claude Code's instruction file to make it act as an expert UK tax adviser. It tells the agent to send out several helpers at once, each covering one area like personal tax or current rates. Each writes notes, and the agent combines them into one answer. He tested it on past professional tax exam questions.
How it went
After about 5 minutes the agent wrote its answer. He gives it 2.5 out of 3 marks, missing one detail about where someone lives for tax. He says plain Opus got the same first question completely wrong.
Worth knowing
The folder doesn't need to be well organised, though he suspects better structure would help. The main barrier is that Claude Code runs in a text-only command window, which puts off non-technical users.
Try it yourself with Claude Code
Write a scraper that recursively downloads the reference documents for [domain, e.g. UK tax] from [source websites] as text files into folders. Then draft a CLAUDE.md that makes you an expert in [domain] who answers questions from non-technical users using only this folder. For each question, launch parallel subagents for [research areas]. Each one writes findings to research_[topic].md, and you consolidate an answer with citations in output.md. Show me the CLAUDE.md for review before we test it on [sample question].
Swap the highlighted [parts] for your own details.
Source: Personal blog · Martin Alderson
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
