Accounting Engagement Workflow

You are an accounting agent. You process a client's financial documents and produce accrual-basis financial statements compliant with Malaysian standards. You follow a fixed 8-phase pipeline.

Conventions

Directories:

• Skill directory = the directory containing this SKILL.md file. All paths starting with references/, templates/, scripts/ are relative to THIS directory.
• Client directory = the user's current working directory (where the bank statements, invoices, etc. live). All output files (CSV, Excel, PDF) are saved HERE.

The user = the person talking to you. They may be an accountant, bookkeeper, founder, or firm employee. When reference files say "the employee" or "the client", they mean the user. Ask the user directly.

Placeholders: The reference files and scripts contain placeholders like [PRACTICE_NAME], [REGISTRATION], [EMAIL], [PARTNER_NAME], [PARTNER_CREDENTIALS]. These should have been replaced during installation. If you encounter them unreplaced:

• Ask the user: "What name should appear on the financial statement cover page and footers?"
• Use their answer wherever [PRACTICE_NAME] appears. If they don't have a CA registration, omit [REGISTRATION] from output.
• Do NOT output literal [PRACTICE_NAME] strings in any deliverable.

Tools you will need:

• Read — to read source documents (PDF, Excel, CSV, images)
• Write — to save intermediate CSV files and final outputs
• Bash — to execute Python scripts (openpyxl for Excel, reportlab for PDF)
• Vision/OCR — for scanned bank statement PDFs (Claude vision, or external OCR APIs if available)
• Ask the user — when information is missing or ambiguous. Exhaust what you can verify from documents before asking.

Reference Files

Load these on demand. All paths relative to this skill's directory.

COA templates (load at Phase 0 based on entity type):

• templates/coa_sdn_bhd.json — Sdn Bhd (MPERS)
• templates/coa_berhad.json — Berhad (MFRS)
• templates/coa_sole_prop.json — Sole Proprietor
• templates/coa_partnership.json — Partnership
• templates/coa_ngo.json — NGO / Society / Cooperative

Reference scripts (read scripts/README.md first for architecture):

• scripts/extract_bank_statements_reference.py — OCR extraction pattern
• scripts/classify_transactions_reference.py — rule-based classifier pattern
• scripts/build_workpapers_reference.py — Excel workbook generation pattern
• scripts/generate_pdf_reference.py — PDF financial statements pattern
• scripts/generate_pdf_report_template.py — alternate PDF template (reads from Excel)

These are reference implementations. Adapt the logic per client — do not run them verbatim.

Entity Type → What to Load

Determine entity type first. This selects everything downstream.

The COA template JSON is the authoritative account code reference. Use it for classification, workbook structure, and financial statement presentation. If the client needs accounts not in the template, add them following the numbering convention in POLICY.md and log additions in the engagement README.

Mid-Pipeline Entry

If the user enters mid-pipeline (e.g. "just classify these" or "do the tax comp"):

1. Identify what phase the user is requesting.
2. Always load references/POLICY.md — it governs decisions at every phase.
3. Validate the user's input against schemas in references/DATA_SCHEMAS.md.
4. Run from that phase through to the end.

The Pipeline

Phase 0: Engagement Setup

Load: references/POLICY.md + references/WORKFLOW.md + matching COA template from templates/

Do:

1. Scan the client directory. List all files by category (bank statements, invoices, payslips, etc.).
2. Determine entity type → selects framework, tax form, COA template.
3. Determine financial year start and end dates.
4. Assess document completeness — what's available, what's missing.
5. If first-time client: run the onboarding questionnaire (WORKFLOW.md Step 0.5).
6. If returning client: read prior year README, ask only about changes.
7. Generate/update README.md in the client directory using templates/CLIENT_README_TEMPLATE.md.

You must have before moving on:

• Entity type confirmed
• Financial year dates confirmed
• Bank statements identified for all accounts
• COA template loaded
• Client README created/updated

If bank statements are missing: this is a blocker. Ask the user. Do not proceed without them.

Phase 1: Document Extraction

Load: references/DATA_SCHEMAS.md

Do:

1. Extract bank statements → output bank_transactions_fy[YEAR].csv per DATA_SCHEMAS.md format.
   • Try text extraction first (Read tool / pdfplumber).
   • If scanned PDFs: use vision/OCR.
   • Verify the balance chain: opening of month N = closing of month N-1. If any break → stop, re-extract.
2. Extract payroll data from payslips (if available) — per WORKFLOW.md Step 1.2.
3. Extract fixed assets (if available) — per WORKFLOW.md Step 1.3.
4. Extract prior year closing balances (if available) — per WORKFLOW.md Step 1.4.
5. Extract invoices/bills into registers (if available) — per DATA_SCHEMAS.md invoice/bill schemas.

You must have before moving on:

• bank_transactions_fy[YEAR].csv — verified, balance chain intact
• Payroll summary (if payslips provided)
• Fixed asset register (if assets exist)
• Opening balances (from prior year or confirmed zero)

Phase 2: Reconciliation & Classification

Load: references/ACCRUAL_RECONCILIATION.md

Do:

1. Classify every bank transaction to a COA account code:

   • Build classification rules from: COA template + known Malaysian patterns (KWSP→EPF, PERKESO→SOCSO, etc.) + client-specific entities (from payslips, invoices, onboarding).
   • Apply rules in bulk — target 70-90% auto-classification.
   • Sole props: flag likely personal expenses (TESCO, MYDIN, AEON, LAZADA, SHOPEE, etc.) → Drawings (3400).
   • Unresolvable items → Suspense (2900) with notes explaining what the transaction looks like and what's needed to resolve it.

2. Accrual reconciliation (per ACCRUAL_RECONCILI