Keep Each Client's Coding Rules in a Claude Project
What the source describes
Anthropic's help center describes Projects as self-contained workspaces with their own chat histories and knowledge bases. Anything uploaded to a project's knowledge base is used across all chats in that project, and project instructions tell Claude how to behave and respond in every chat there. Projects are available on free accounts, which can create up to five. Context is not shared across chats within a project unless it is added to the project knowledge, and on paid plans Claude automatically enables RAG mode when project knowledge approaches the context window limit.
Result or outcome
According to the source, the project knowledge base sits on the right side of the project's main page, and Claude processes the documents, text files or snippets you upload there and uses them as context in every chat within the project.
How to run it
- Create one project per client at claude.ai/projects, named so the client is obvious to everyone on your team.
- Upload the client's chart of accounts, a coding rules document and a few months of reviewed transactions to the project knowledge, with account numbers and personal details removed.
- Click Set project instructions and paste the prompt on this page, filled in for that client.
- Each month, start a new chat inside the client's project and paste or attach the uncoded transactions.
- When you approve a new rule, add it to the rules document in project knowledge, because chats in a project do not share context otherwise.
Try it yourself
Copyable prompt
You help me categorize bank and card transactions for [client name, business type]. Always use the chart of accounts and the coding rules in this project's knowledge, and never an account that is not on that chart. For each transaction I paste, return: date, description, amount, suggested account, the rule or past example you relied on, and a confidence of high, medium or low. Mark anything that could be personal, a transfer between accounts, a loan payment or an owner draw as NEEDS REVIEW, and do not guess. If no rule fits, say so and suggest a new rule for me to approve. Never decide tax treatment. Return a table I can paste into a spreadsheet.
Replace the [bracketed] placeholders. Remove client names, account numbers, and personal data before pasting into any AI tool.
What to review
Spot-check suggested accounts against source documents, especially low-confidence and NEEDS REVIEW lines, before you post anything. Keep the rules document current, since Claude uses whatever is in project knowledge even after a rule changes. On Team plans, check each project's visibility so client data stays with the right people.
Related free download: 50 AI Prompts for Bookkeepers (Prompt pack (PDF))
Source: Claude Help Center (What are projects?); also Claude Help Center (How can I create and manage projects?)
Not professional advice. This is general information, not accounting, tax, legal or financial advice. Check AI output yourself and consult a licensed professional for your situation. See our Terms of Use.
This entry summarizes a public source. Features, plans, and availability change, so confirm details with the vendor. Tool names are trademarks of their respective owners.
Related workflows
- Receipts and Expenses
Categorize Leftover Expenses with AI After Bank Rules
This Brocent guide argues that QuickBooks and Xero bank rules should handle recurring transactions deterministically, and that ChatGPT is best used for the residue of unrecognised payees, split transactions and anomaly review. It recommends exporting uncategorised transactions to CSV with the real chart of accounts, asking for a suggested category with a confidence note and flags, and reviewing everything before posting. It stresses never auto-posting AI output, keeping an audit trail and leaving tax judgements to an accountant.
Works in ChatGPT, Claude, Grok, Gemini
QuickBooks Online, Xero
Brocent Blog - Practice Operations
Clean Up a Messy Chart of Accounts with an AI Review Plan
OCTA's resource page describes AI chart of accounts cleanup that reviews a client's account list against actual posting activity to find duplicates, near-duplicates, unused accounts, wrong account types, misparented sub-accounts and naming inconsistencies. In OCTA Flow the agent connects to QuickBooks, Xero, Sage or Zoho, proposes merges, reclasses, deactivations and renames ranked by severity, and the bookkeeper applies approved changes in the ledger. The page notes that merges can move history and affect prior-period reports, so those are flagged for close review.
OCTA Flow, QuickBooks Online +3
OCTA (vendor resource) - Reconciliations
Reconcile a bank statement export against your chart of accounts with Claude Cowork
Claude Academy walks through giving Cowork a folder with a checking account statement export and the chart of accounts, then asking it to match transactions to accounts, flag anything that does not line up, and draft journal entries for discrepancies. Claude may ask clarifying questions and build a plan for review, then returns the reconciliation logic, specific discrepancies and journal entries ready for review. Suggested follow-ups include exporting entries as CSV and creating a reconciliation summary for the records.
Claude, Claude Cowork +1
Claude Academy (Anthropic tutorial) - Receipts and Expenses
Chase Clients About Uncoded Transactions with an AI-Drafted Email
The Firm's guide to ChatGPT prompts for bookkeepers argues that bank feeds and ledger rules already handle matching, so chat assistants are most useful for the writing around the ledger, starting with the email asking a client to explain transactions the bookkeeper cannot code. It recommends grouping similar items, numbering questions, requesting receipts, and following up with a shorter, firmer second chase. The article also says to strip names, tax identifiers and account numbers before pasting, use a firm-approved tier, and never let the assistant decide tax treatment or code from memory.
Works in ChatGPT, Claude, Grok, Gemini
The Firm (accounting publication)