How to Write Better AI Prompts for Accounting and Tax Work

Better AI prompts for accounting and tax work are less about clever wording and more about forcing Indian context, naming the deliverable, and blocking unsafe shortcuts. A weak prompt (“Explain GST ITC”) gets a blog-post answer. A strong prompt (“AY 2025-26, trading client, list GSTR-2B vs books checks; mark portal confirmations; do not invent circulars”) gets something a reviewer can use. This is the prompt craft practising CAs need — complementary to how to use ChatGPT for CAs and the ready-made 50 prompts list.

The five-part prompt skeleton

Use this almost every time:

1. Role — “Act as an Indian CA reviewer / tax manager / GST specialist.”
2. Context — AY or tax period, assessee type, industry, GSTIN count if relevant (anonymised).
3. Task — checklist / outline / rewrite / competing arguments / data fields needed.
4. Constraints — “Do not invent section numbers or case names”; “Do not change any figure”; “Leave rates as [VERIFY].”
5. Output shape — bullets, table, SOP columns, or “questions only.”

Skip any of these and the model fills gaps with confident fiction.

Patterns that work in tax and accounting

1. Checklist prompts

Ask for checks, not conclusions.

“Draft a pre-filing GSTR-3B checklist for a multi-GSTIN manufacturer. Separate data-entry checks from judgement calls. No fabricated notifications.”

2. Competing-arguments prompts

For grey areas:

“List arguments for treating [anonymised facts] as capital vs revenue for AY [YYYY-YY]. Include factors an Assessing Officer will weigh. Do not pick a winner. No invented cases.”

3. Rewrite-without-changing-numbers

“Rewrite this verified computation schedule for a non-finance director. Do not alter any number. Flag sentences that still need CA confirmation.”

4. Placeholder law

“Outline a reply structure for an anonymised ITC mismatch SCN. Use [CITATION NEEDED] wherever law would appear. Do not invent circulars.”

5. Data-fields-first

Stop the model inventing figures:

“Before any computation, list the data fields you need from AIS, 26AS, and books for [anonymised capital-gains facts]. Do not compute yet.”

Anti-patterns (stop doing these)

| Weak prompt | What goes wrong | |---|---| | “Is this deductible?” | False certainty on judgement | | “Cite the relevant cases” | Invented authorities | | “Calculate the tax” with incomplete facts | Assumed numbers that look finished | | Paste full ledger + PAN into consumer chat | Confidentiality / DPDP exposure | | “Update me on latest GST law” with no date | Stale or blended rates |

Failure modes in depth: what AI gets wrong on Indian tax. Data rules: DPDP and AI tools.

Prompting domain tools vs ChatGPT

On ChatGPT / Claude / Gemini, prompts carry almost all the India specificity. On Taxmann.ai / TaxBotGPT / VIDUR / Vaive, the corpus helps — but you still specify AY, facts, and “show sources,” then open them. On GSTAgent-class recon tools, the “prompt” is often configuration and exception review, not a chat box — do not pretend a paragraph of English replaces matching logic (GST recon with AI).

Make prompting a firm habit

• Maintain a shared prompt library with owners (GST / IT / audit).
• Require anonymisation in the library header.
• Teach juniors: draft → verify → file trail.
• Review prompts in training the same way you review working papers.
• Escalate when output invents a section number — that is a quality event, not a joke.

For notice-specific prompting, see prompting AI for GST notices. Model choice among general LLMs: ChatGPT vs Gemini vs Claude for CAs.

Worked example: from weak to strong

Weak: “How do I reply to a GST notice for ITC mismatch?”

Strong: “Act as an Indian GST practitioner. Context: anonymised trading client, FY 2024-25, theme is GSTR-2B vs books ITC mismatch on B2B inward supplies. Task: outline the sections a reply usually contains and a checklist of evidence to gather from books and portal. Constraints: do not invent circulars, case names, or section numbers — use [CITATION NEEDED]. Do not draft a final legal position. Output: numbered outline + evidence table columns.”

The strong version tells the model what not to do, which is how you keep juniors from pasting fictional CBIC references into Word.

Prompt chaining (without losing the plot)

1. Structure — outline only.
2. Evidence list — what to pull from Tally / Excel / portal.
3. Draft prose — only after evidence exists, with numbers you supply.
4. Client plain-English — rewrite of the verified draft.
5. Reviewer questions — ask the model to attack the draft; you decide what sticks.

Chaining beats one mega-prompt that tries to research, compute, and persuade in a single breath.

Sector-specific add-ons

Add two lines when the industry changes the risk:

• Real estate / works contracts: remind place-of-supply and credit restrictions as questions to verify, not answers to invent.
• Exporters: separate LUT / refund evidence from generic ITC talk.
• Professionals under presumptive schemes: force the model to ask whether 44AD/ADA-style facts apply before drafting ITR logic.

Measuring prompt quality in review

Track for a month: (a) drafts rejected for invented citations, (b) drafts rejected for wrong AY, (c) drafts accepted with only stylistic edits. If (a) stays high, your library constraints are too soft. If (c) rises, promote those prompts into the firm SOP and retire the vague ones.

Frequently asked questions

What makes a good AI prompt for tax work?

A good prompt states the Indian context (AY, assessee type, GST vs IT), the task type (checklist, competing arguments, rewrite), and hard constraints (do not invent citations; do not change numbers). Vague prompts produce fluent, US-flavoured answers you cannot defend.

Can better prompts eliminate AI hallucinations in tax?

No. Better prompts reduce how often you copy a fake section into a Word file. They do not make a general model a primary-law database. You still open provisions and portal extracts before client delivery.

Should prompts ask the AI to decide capital vs revenue?

Prefer asking for competing arguments and decision factors. Asking for a single verdict invites false confidence on judgement calls that remain the CA’s responsibility.

Where should I put example prompts for my team?

Keep an approved prompt library next to your AI policy — GST, income tax, audit, Excel, client email — with anonymisation rules. A starting set is in our 50 ChatGPT prompts for Indian CAs guide.