TallyPrime Entry Skill

Connect to a locally running TallyPrime instance via XML-over-HTTP. All requests are HTTP POST to $TALLY_URL (commonly http://localhost:9000).

Scope

This skill handles only:

For anything outside this scope, tell the user this skill does not cover that task.

Critical rules (must follow)

1. Never assume company — if not explicit, ask which company before posting.
2. Never guess ledgers — verify ledger exists before posting; ask user if unsure.
3. Dates are YYYYMMDD (no separators).
4. Unique GUID per voucher — prevents duplicates on retries.
5. Voucher totals must balance — sum of all AMOUNT values = 0.
6. Escape XML — & → & in narration and party names.
7. Confirm before posting — every Create is a write; confirm company and intent first.
8. Bill-wise allocations on party ledgers — include BILLALLOCATIONS.LIST to keep outstandings correct.
9. LEDGERENTRIES.LIST in Invoice Voucher View — never use ALLLEDGERENTRIES.LIST when OBJVIEW="Invoice Voucher View" is set; Tally silently ignores it.
10. Post-entry review is mandatory — after every voucher post, fetch it back and verify before telling the user it is done.
11. Voucher class — confirm before posting — run the class decision rules at the top of the relevant purchase/sales reference file before posting.
12. Check import response — after every POST, read CREATED/ERRORS from Tally's response (see reference/reports.md). Never confirm success if ERRORS > 0 or CREATED = 0.

Step 0: Server check

curl -s --max-time 5 "$TALLY_URL"

Expected response:

\x3CRESPONSE>TallyPrime Server is Running\x3C/RESPONSE>

If no response, ask the user to open TallyPrime and enable integrations on the correct port.

Step 1: Company

Ask for the company name if not provided. Use the exact spelling in SVCURRENTCOMPANY.

Step 2: Invoice / image extraction (mandatory OCR)

When the user sends a PDF, JPEG, JPG, or PNG invoice or bill, use the scribe CLI. Do not extract data with the LLM directly.

PDF extraction

mkdir -p "scribe-extractions"
scribe type "\x3Cinput.pdf>"
scribe extract -f txt "\x3Cinput.pdf>" "scribe-extractions/\x3Cid>.txt"

Read the .txt file and parse the invoice fields. If quality is poor (image-native PDF), run OCR first:

scribe recognize -h -o "scribe-extractions/\x3Cid>-ocr" "\x3Cinput.pdf>"
scribe extract -f txt "scribe-extractions/\x3Cid>-ocr/\x3Cstem>.pdf" "scribe-extractions/\x3Cid>-recognized.txt"

Image extraction (JPEG / PNG)

1. Try direct extraction:

   scribe extract -f txt "\x3Cinput.jpeg>" "scribe-extractions/\x3Cid>.txt"
   

2. If output is empty, run OCR:

   scribe recognize -h -o "scribe-extractions/\x3Cid>-ocr" "\x3Cinput.jpeg>"
   scribe extract -f txt "scribe-extractions/\x3Cid>-ocr/\x3Cstem>.pdf" "scribe-extractions/\x3Cid>-recognized.txt"
   

After extraction

Present a clean summary to the user: supplier/customer, GSTINs, invoice number, date, line items, taxable value, CGST/SGST/IGST, round-off, total. Ask for confirmation before posting.

If extraction fails, say "I could not read this bill clearly. Please resend a clearer photo or PDF." — do not mention CLI, OCR, errors, or paths.

Step 2b: Verify / create masters

Before posting any voucher, verify that all required ledgers and (for Item Invoice) stock items exist. Use reference/ledgers.md and reference/inventory.md.

Ledger check (reference/ledgers.md):

• Fetch List of Accounts and confirm party ledger, purchase/sales ledger, and GST ledgers exist.
• If any are missing, create them using the templates in reference/ledgers.md before posting.

Inventory check — Item Invoice only (reference/inventory.md):

• Confirm each stock item referenced in the invoice exists in Tally (fetch Stock Summary).
• If missing, create in order: Stock Group → UOM → Stock Item.
• Confirm godown: default is Main Location / Primary Batch unless the company has custom godowns.

Bank statement pre-flight (reference/reports.md):

• Before posting bank entries, fetch all ledger names (Ledger Names query in reference/reports.md).
• Present the mapping to the user: "These are the ledgers I will use: [bank], [parties]. Confirm to proceed."
• Do not post until confirmed.

Step 3: Post vouchers

Load only the file that matches the task — this keeps context lean:

Voucher class decision — the class decision rules and GST ledger name table are at the top of each purchase/sales reference file. Run them before posting.

Preflight checklist before any Create:

Step 4: Post-entry review (mandatory)

After every voucher Create:

1. Read the import response — check CREATED and ERRORS fields (reference/reports.md). Stop if ERRORS > 0.
2. Fetch the voucher back — use Voucher Register (preferred) or Day Book as fallback (reference/reports.md).
3. Match against the intended entry — company, voucher type, number, date, party, all ledgers, amounts, GST split, bill-wise allocations.
4. If everything matches, summarize to the user in accountant language (no XML, no HTTP terms).
5. If anything is off, do not confirm completion — explain the mismatch and ask how to proceed.

Never say "Entry posted" without completing this review.

PDF Generation — Sales Invoice

Use the tallyca CLI. Do not build PDFs manually.

Preflight

tallyca --version   # must be >= 1.1.0

If missing or older: npm install -g tallyca@latest

Minimum user input

Ask the user for only these 4 fields. Everything else is auto-filled from Tally:

Auto-fill sequence (run before calling tallyca)

Execute in this order. Each step reduces what the user needs to provide:

Fallback: if any auto-fill fails, use --json-errors. The CLI returns a missing array — ask the user only for those specific fields, then retry.

From structured fields (preferred)

Always use --no-interactive and --json-errors when running non-interactively:

tallyca generate:invoice \
  --company "ABC Company" \
  --company-gstin "GSTIN_HERE" \
  --company-address "Full company address, State PIN" \
  --party-name "XYZ Party" \
  --customer-gstin "PARTY_GSTIN" \
  --billing-address "Party address, State" \
  --invoice-no 186 \
  --date "2/1/2026" \
  --place-of-supply "Uttar Pradesh" \
  --item "PQR Item" \
  --qty 140 \
  --unit Bag \
  --rate 279.66 \
  --hsn-code 25322210 \
  --gst-rate 18 \
  --b2b \
  --output invoice_186.pdf \
  --no-interactive \
  --json-errors

On exit code 2 (validation failure): parse JSON from stderr, read the missing array, ask the user only for those fields, then retry. Never hand-build a PDF.

{"error":"validation","message":"Missing required invoice fields","missing":["place-of-supply","unit"],"warnings":[]}

From raw WhatsApp / Telegram text

tallyca from-text \
  --company "ABC Company" \
  --text "Party Name: XYZ Party
Invoice No.: 186
Date: 2/1/2026
Place of Supply: Uttar Pradesh
Item: PQR Item 2523 @ 18 %
Qty: 140 Bag
Rate: 279.66/Bag
HSN Code: 25322210
Amount: 39152.40
Make sure to use voucher class Sales @ 18 %" \
  --output invoice_186.pdf \
  --no-interactive \
  --json-errors

On Linux / AWS without Chromium

export TALLYCA_PDF_BACKEND=pdfmake

Then run the same command. Do not substitute a hand-built PDF.

GUID pattern

{companyShort}-{voucherType}-{invoiceNo}-{date}

Examples: abc-purchase-inv001-20260115 · abc-sales-186-20260201

Accountant-friendly responses

After posting, do not mention XML, HTTP, payloads, status codes, or integration internals. Say:

• "Entry posted" (not "XML import succeeded")
• "Tally is connected" (not "server responded 200")
• Summarize: company, voucher type, date, party/ledger names, amounts, tax split, narration.