CPA Firm Engine for Claude
Give every client workstream a clear route to review
Bring inquiries, records, bookkeeping questions and review notes into one coordinated drafting workflow. Use eight specialist role prompts with a shared router to prepare useful internal work, preserve the evidence behind it and hand the next decision to the right person.
This downloadable kit gives you a practical starting point in Claude: editable instructions, a complete synthetic case and a shared review queue you can inspect before adapting the process to your firm.
One shared context. Eight focused roles.
Supply authorized QuickBooks exports, Gusto payroll exports, Gmail inquiry or notice extracts, and Google Drive documents. The router starts with your firm SOPs, confirms the source inventory and routes work according to prerequisites. Each role has detailed steps, output requirements, escalation gates and a queue handoff.
Intake
Turn a new inquiry into a scope matrix, engagement outline and confirmation questions. Separate requested work from agreed services, and keep missing entity, jurisdiction and acceptance details visible.
Document chaser
Build one consolidated missing-record list and a draft follow-up. Specify the document, period, reason and owner, while avoiding duplicate requests and keeping delivery inside the firm’s review process.
Bank coding
Suggest categories from an approved historical merchant map. Preserve transaction IDs and amounts, explain the match, and flag unfamiliar merchants or unsupported business purposes for a human decision.
Month-end close
Prepare a provisional cash reconciliation, payroll cash comparison and open close checklist. Show the calculation and unresolved differences instead of treating an unexplained balance as a completed close.
Return review
Cross-check supplied draft fields against same-year source workpapers. Highlight mismatches, missing support and limitations, with review notes for the CPA rather than unsupported corrections or filing actions.
Notice desk
Extract notice details, surface urgent verification questions and prepare a reply outline for review. Keep stated dates separate from verified deadlines, and flag missing pages or representation authority.
Planning scout
Convert supported client intentions into focused CPA questions. Identify missing facts and dependencies without inventing tax savings, eligibility conclusions or recommendations that the evidence cannot support.
Partner pack
Combine the specialist drafts into a decision-focused client-status brief and meeting agenda. Reconcile queue counts, retain blockers and assign clear owners, actions and internal review checkpoints.
Working files you can inspect and adapt
- Eight agent prompts, ROUTER.md and CLAUDE.md define the coordinated workflow, shared rules and human approval boundary.
- Eight blank templates cover client context, source inventory, transactions, historical categories, bank summary, payroll, return-field comparisons and the review queue.
- A fully synthetic case and eight authored draft examples show the expected depth, evidence references and handoffs. The example includes deliberate exceptions for reviewers to investigate.
- An offline HTML guide and review board make the files easy to explore. The optional standard-library Python helper checks schemas, evidence references, periods and exact decimal comparisons, then rebuilds the board.
Keep the evidence attached to the decision
Every queue row carries a draft version, evidence IDs, confidence, owner, next action and internal checkpoint. Source IDs resolve to specific records or document passages. Dependencies keep unresolved work visible when another role summarizes it.
The supplied queue contains eight unapproved drafts. The workflow distinguishes draft, blocked, ready for review and human-approved work; an assistant cannot grant approval. A material revision returns to review. Matching arithmetic confirms only the stated comparison, never the completeness of the books or correctness of a return.
Start with the example, then use your own records
Open the offline guide.
Extract the entire ZIP and double-click README.html. Follow its clickable file links, inspect the synthetic case and open the included review board.
Upload the working instructions.
In a Claude chat with file attachments, supply CLAUDE.md, ROUTER.md, the context, selected agent prompts and source files. Use the guide’s copyable kickoff prompt; stage uploads if needed.
Prepare coordinated drafts.
Ask Claude to confirm accessible sources and missing prerequisites, then apply the role instructions in dependency order. Keep the current queue and evidence snapshot together throughout the session.
Review, then adapt.
Compare your training results with the eight worked drafts. For real work, complete the blank templates, replace the baseline SOP with approved procedures and assign named human reviewers before using client data.
Built for teams preparing recurring client work
Use the kit when several people or workstreams need a consistent way to prepare drafts, request evidence and surface decisions. Adapt the prompts to your service scope, source formats and review responsibilities; retain the shared evidence and approval discipline.
A guided workflow with a defined review boundary
The agents are coordinated role prompts, not deployed background software. You provide the files or independently configured, approved connectors. The ZIP does not set up account integrations or send messages, post entries, sign engagements, file returns or make payments.
Claude access is needed to use the prompts interactively. The guide and sample board open locally without a server. Python 3.10 or later is optional for the local helper, which uses no AI, network or credentials. Professional judgments and outward actions remain in your firm’s authorized human process.