AI prompts for tax research memos (with citation-verification built in)

Why even grounded tax AI hallucinates

Three failure modes recur across every tool on the market.

• Stale corpus. A Revenue Ruling revoked by later guidance still returns as “good law” because the obsoletion notice has not propagated to the index the model is querying.
• Inferred citations. Asked to support a position with primary authority, a model would rather generate a plausible-looking cite than admit no on-point authority exists. The cite looks like one you have read a hundred times. That is the entire point.
• Misattribution. A passage from a treatise, a Chief Counsel Advice, or a Private Letter Ruling appears in the draft as if it were binding. The reasoning is real. The weight is wrong. You cite a memo as if it were a regulation, and the position collapses on first contact with review.

Grounded tools narrow these failure modes. They do not close them.

The four checks every citation must pass

Burn this into muscle memory. Every citation in every AI-touched memo must clear four gates:

1. Existence — the document is real at the cited identifier.
2. Pinpoint — the cited subsection, paragraph, or page actually contains the proposition.
3. Currency — the authority has not been revoked, superseded, overruled, modified, or withdrawn.
4. Weight — primary binding, primary non-binding, persuasive, or commentary.

The prompts below build all four into the request so the model cannot return output that skips any of them.

The six core prompts

Each prompt uses the same shape: role, task, corpus, output format, verification block. Replace the bracketed parameters with your jurisdiction and source set.

1. Issue spotter

Role: senior tax associate.
Task: extract every tax issue raised by the fact pattern below.
Fact pattern: [paste facts]
Jurisdiction: [e.g., U.S. federal income tax]
Output:
- Numbered list of issues, each <= 20 words.
- For each: the body of law it implicates (statute, regulation, doctrine).
- Separate list: issues you considered but excluded, with one-sentence reasons.
Verification: do not assert any authority yet. If an issue cannot be mapped to a specific body of law, mark it "open — research required."

2. Authority research

Role: tax research librarian.
Task: return primary authorities on issue [X] in [jurisdiction].
Corpus: only the documents available in [Bloomberg Tax / Checkpoint Edge / CCH / uploaded PDFs]. Do not draw on outside sources.
Output (table):
Cite | Type (statute / reg / ruling / case / commentary) | Holding (<= 25 words) | Weight (binding / persuasive / commentary)
Verification:
- If no on-point primary authority exists in the corpus, return "no primary authority found" — do not substitute commentary.
- Quote the exact pinpoint language for each holding.
- Flag any authority older than five years for currency check.

3. Counter-authority

Task: find the strongest authority that cuts against the position "[state position in one sentence]."
Corpus: [same as above].
Output:
- Ranked list (most damaging first).
- For each: cite, holding, why it hurts, and the cleanest distinguishing argument.
Verification: same table format and currency flag as the research prompt. Mark "no contrary authority found in corpus" if true — do not invent counter-authority for symmetry.

4. IRAC memo draft

Task: draft the analysis section using IRAC.
Constraint: cite only authorities from the approved list below. Reject any cite not on this list, even if you "know" it.
Approved authorities: [paste the cite column from Prompts 2 and 3]
Output:
- Issue, Rule, Application, Conclusion for each issue.
- After every cite, insert [pinpoint: §X(a)(2)] or the equivalent.
Verification: at the end, list every cite used in the draft and confirm each appears on the approved list. If any do not, stop and report.

5. Citation audit

Task: extract every citation in the memo below.
Memo: [paste draft]
Output (table):
Cite | Proposition supported | Pinpoint quoted | Weight | Source confidence (high / medium / low)
Verification:
- Mark "low" any cite that cannot be matched to a specific document in the corpus.
- Mark "medium" any cite where the pinpoint language does not directly support the proposition.
- Flag any cite that appears to be commentary presented as primary authority.

6. Currency check

Task: for each cite below, return the most recent action affecting its status.
Cites: [paste from audit]
Output (per cite):
- Status: in force / revoked / superseded / modified / overruled / distinguished / withdrawn.
- Date of the action and the superseding authority, if any.
- If you cannot verify currency from the corpus, mark "currency unverifiable — manual check required." Do not guess.

The 10-minute ritual

For every memo with AI involvement, run this before sending:

1. Run Prompt 5 on your own draft.
2. Run Prompt 6 on whatever survives.
3. Open every cite still standing. Confirm existence and pinpoint in the actual source — Bloomberg, Checkpoint, CCH, Westlaw, or the issuing authority’s own website.
4. Save the verified citation list alongside the memo file.

One minute per cite, four checks each. A ten-cite memo costs you ten minutes and removes the failure mode that ends careers.