Why a Japanese Accounting Firm’s Most Critical “Employee” Is a Markdown File

Overview

A Japanese tax accountant with no programming background runs a firm for 60 clients and leaves the office at 5 pm. He uses a single Markdown file named CLAUDE.md that Claude Code reads each night to process unhandled accounting entries automatically.

Technical analysis of CLAUDE.md

CLAUDE.md

is a persistent instruction file that Claude reads before any work. It acts as an AI employee handbook, defining the AI’s role, permitted actions, tool interactions, and data‑handling policies. Claude Code supports user‑level, project‑level, and organization‑level files; higher‑priority files override lower‑priority ones. Large rule sets can be split into .claude/rules/ and imported with @path. Files are recommended to stay under roughly 200 lines to limit context consumption. The special command /init can generate an initial CLAUDE.md.

Four‑step construction

Step 1: Define the AI workspace

Before writing any instructions, a folder hierarchy is created so the AI knows where client files, logs, tools, and reference documents reside:

ai-management/
├── output/      # AI‑generated files
├── reference/   # Documents AI can read
├── clients/     # Per‑client folders
├── tools/       # Automation scripts
├── work‑log/    # Logs
├── finance/     # Accounting‑related

Step 2: Write CLAUDE.md

The file is organized into four modules that turn the workflow into an AI‑executable system.

Module 1 – Define the AI role

“You are the executive layer of an AI tax firm.” The AI handles execution (book‑keeping, drafting documents, task organization) while humans retain judgment (tax decisions, special cases, final approval).

Module 2 – Set red‑line boundaries

“AI must never make tax judgments.” Known patterns are routed to the AI; ambiguous or regulated cases are routed to humans; autonomous booking is blocked for specific transaction types.

Automatically excluded from autonomous booking:
1. Unclear charges
2. Loan repayments
3. Social insurance and taxes
4. Payroll
5. Investment activity
6. ATM withdrawals
7. Utilities requiring special handling

Module 3 – Connect tools

- Use freee to retrieve unprocessed transactions
- Register journal entries only after classification is confirmed
- Use Google Calendar to check today’s schedule before generating tasks
- Use Notion for storing meeting notes and TODOs
- Use Gmail only for drafting replies, not sending without approval

Module 4 – Safety design

Risk principles are encoded so that before any action the AI checks security rules, decides whether to proceed, mask data, or report, making safety part of the decision loop.

Night‑time accounting run

At 21:00 Claude starts a batch process in freee, retrieves unprocessed transactions, normalizes their descriptions, and classifies them against a keyword dictionary. Deterministic matches are processed automatically. Ambiguous cases are handed to a fallback classifier that uses masked data to infer the most likely category. High‑confidence results are written back; low‑confidence cases are left for human review after a duplicate check. The workflow therefore follows a “selective autonomy” pattern: routine cases are handled by the AI, while new, ambiguous, or high‑risk cases require human judgment.