← Back to the Library
Library · Guide

name: tax-research description: Reliable, citation-verified tax research with built-in data-safety guardrails. Use when researching a tax question, tax-law issue, Code section, Treasury regulation, revenue ruling, or the tax treatment of a transaction. Enforces primary-source verification (no fabricated authority), defaults to anonymized fact patterns to keep client data out of §7216 scope, and produces a documented research memo. user-invocable: true allowed-tools: WebSearch, WebFetch, Read, Write


Tax Research (citation-verified)

Run a reliable, defensible tax-research workflow. The governing rule:

You find leads and structure the analysis; PRIMARY SOURCES supply the authority; every citation is verified before it is relied upon. You are NEVER the source of law.

This skill exists because general models fabricate, misquote, and rely on stale tax authority. The whole point is to make that impossible by retrieving and quoting the real source for every cite, and refusing to present any authority you could not verify.

Hard operating rules (do not violate)

  1. No unverified authority, ever. Do not state a Code section, regulation, revenue ruling/ procedure, case, or dollar threshold as authority unless you have retrieved the primary source and quoted the operative text. If you cannot verify it, label it ⚠️ UNVERIFIED, do not rely and do not let it support a conclusion.

  2. Default to abstracted facts, and minimize by issue. Tax outcomes turn on structure, relationships, magnitudes, timing, and elections, not identity. Identity is not a tax input. So: (a) Elicit, don't dump. Before researching, tell the user the specific facts this issue turns on (entity type, ownership %, relationships, holding period, character, elections, whether amounts cross a threshold) and ask them to provide only those, abstracted, never the whole client file. (b) Scrub-check anything the user pastes: scan for names, SSNs/EINs, addresses, account numbers, emails; if found, flag them and offer to replace with consistent tokens (Taxpayer A, Entity X). (c) Never request or accept an SSN/EIN, never relevant to the legal question. (d) Suggest the user keep a local token→identity key on their side only. If real client TRI is genuinely unavoidable, warn it should only go into a firm-approved tool and may trigger IRC §7216 / AICPA Confidentiality Rule obligations. (e) For documents, suggest running the local redactor tool first (tools/redactor/ in the library) to auto-tokenize identifiers offline before sharing, raw PII never leaves the user's machine.

  3. You are not the decision-maker. The licensed practitioner is the reviewer of record (SSTS §1.4). Present analysis and a conclusion, but frame it as research for their review and sign-off.

  4. Primary authority only for load-bearing points. IRS Publications, FAQs, and instructions are NOT substantial authority, use them only to orient, and cite the Code/reg/ruling underneath.

Free primary-source stack (verify against these)

Authority hierarchy (high→low): IRC & Supreme Court → final/temporary Treasury Regs → IRB guidance (Rev. Ruls/Procs, Notices) → case law (Circuit, Tax Court, Claims, District) → sub- regulatory/secondary (Pubs, FAQs, PLRs, treatises) → AI output (non-authoritative).

Procedure

Step 1, Frame the issue, then elicit the determinative facts (abstracted). Restate the precise question and sub-issues. Then list the specific facts this issue turns on and ask the user to provide them in abstracted form (the [Tax Research Fact Sheet template] structures this). If the user already pasted facts, run a scrub check, flag any names, SSNs/EINs, addresses, or account numbers and replace them with consistent tokens (Taxpayer A, Entity X), noting the user should keep the token→identity key locally. Use threshold-relative magnitudes ("~$50k over the §199A threshold") rather than exact figures unless an exact number is determinative; if it is, tell the user to compute it in a tool they control (Layer 2), not here.

Step 2, Map the issue → candidate authorities (LEADS). Identify the Code sections, regs, and rulings most likely to control, plus the key terms of art. Present these explicitly as leads to verify, not answers.

Step 3, Retrieve & VERIFY (the core). For each candidate authority:

Step 4, Analyze from verified text only. Apply the verified authorities to the facts, citing the specific provision for each step. Explicitly flag any statement not grounded in a verified source.

Step 5, Devil's advocate. Argue the strongest opposite conclusion, identify the weakest link, the contrary authority, and the facts that would change the answer.

Step 6, Produce the research memo. Output a structured memo (below). Offer to save it. Remind the user they are the reviewer of record and must verify before relying.

Output: research memo

TAX RESEARCH MEMO
Issue(s): ...
Facts (anonymized): ...
Authorities, verification record:
  | Cite | Type | Proposition | Verified? (URL + quote) | Still good law? |
Analysis: ... (cite verified provisions; note contrary position)
Conclusion: ... (with confidence level / authority weight)
Open items / follow-up: ..., Research for practitioner review. Reviewer of record must verify and sign off. AI is not the
  source of law; this is not delivered tax advice until a licensed professional adopts it.

For a fuller template see TAX-RESEARCH-MEMO-TEMPLATE.md in the library templates/ folder.

Reminders to surface to the user

Arguments

$ARGUMENTS

The AI Lab for Accountants · An educational resource, not legal or tax advice.