Tax

Classification Workflow

1. Normalize the source data.

   • Inspect encoding, headers, duplicated columns, currencies, and date range.
   • Filter to the target year.
   • Preserve source row IDs if available.

2. Build candidate sets.

   • Split rows into clear include, clear exclude, and manual review.
   • Use deterministic rules first: merchant, category, currency, date, memo, amount pattern.
   • Treat ambiguous rows as manual review rather than forcing a classification.

3. Confirm account mapping.

   • Default expense entry pattern for personally paid business expenses is:
     • 借方 = expense account
     • 貸方 = 事業主借
   • Common mappings:
     • cafe/workspace: 雑費
     • telecom/SIM: 通信費
     • transport: 旅費交通費
     • lodging: do not assume; confirm the user's policy

4. Verify tax-sensitive claims.

   • For legal or tax classification questions, prefer official NTA guidance and primary documentation.
   • If the user wants a nonstandard treatment, follow their instruction for bookkeeping work but distinguish it from general tax guidance.

MoneyForward Automation Workflow

Prefer this order:

1. Use an isolated Playwright-controlled Chrome profile.
2. Let the user sign in manually in that isolated browser.
3. Inspect the actual MoneyForward form fields and select values before bulk posting.
4. Submit one stable accounting pattern at a time.

Avoid:

• reusing encrypted browser state unless explicitly approved and necessary
• depending on the user's live Chrome profile by default
• relying on visual clicks when direct form fields can be identified safely

Verification Workflow

After posting rows:

1. Search the journal list using merchant or remark.
2. Match on:
   • date
   • amount
   • merchant or remark
   • expected account text
3. Record the MoneyForward transaction number for each matched row.

Do not assume success from:

• form reset
• HTTP 200
• absence of an error banner

If verification is inconsistent:

• run a second search pass
• then run a count-based audit by source row
• only retry rows that remain genuinely missing

Reconciliation And Cleanup

If retries created duplicates:

1. Determine the intended count from the source export.
2. Search the MoneyForward journal list by merchant.
3. Identify duplicates by exact row details and transaction number.
4. Delete only the surplus rows.
5. Re-audit until the MoneyForward count matches the source count.

Use transaction numbers for deletion decisions. Do not dedupe by text similarity alone.

Outputs

When the task finishes, provide:

• filtered or normalized source file paths, if created
• imported count
• skipped count
• manual-review count
• duplicate cleanup summary, if any
• list of unresolved rows requiring user judgment

Notes

• Keep the main skill procedural and reusable.
• Put one-off merchant heuristics or session-specific examples in reference files if this skill grows.