Every advisor who has ever opened ChatGPT, Claude, Microsoft Copilot, or Gemini for the first time has typed something like "summarize this tax return" — and gotten back a generic, hedged, vaguely competent paragraph that would embarrass a paraplanner. The output was not the model's fault. The output was the prompt's fault. A prompt for an advisor task is not a question; it is a specification — a five-part contract that tells the model who it is, what it knows, what to do, how to shape the answer, and what it must never do. This lesson installs the five-part anatomy, demonstrates it on the canonical Holistiplan-to-client-one-pager rewrite, contrasts a beginner prompt against a senior advisor's prompt on identical input, and walks the draft-critique-regenerate loop that turns a brittle one-off prompt into a versioned, dated, tested-with-N-clients artifact that lives in the firm's library and survives a Reg BI exam.

Why Every Prompt Is a Reg BI Artifact in 2026

Before the anatomy, the regulatory framing. By May 2026, the operating consensus across the Smarsh, Global Relay, Sidley, Debevoise, and ACA Group practitioner notes on the FINRA 2026 Annual Regulatory Oversight Report and the SEC Division of Examinations Marketing Rule risk alerts is that the prompt is a record. When an advisor types "draft a Reg BI rollover memo for the Hendersons rolling a $480,000 401(k) at Fidelity NetBenefits to a Schwab IRA, with the four alternatives documented and a fee comparison," the prompt becomes part of the artifact chain that downstream produces a recommendation. Under SEC Rule 204-2 the firm must retain the records used to produce advertisements and certain advisory communications; under FINRA Rule 4511 the BD must retain books and records including the inputs that produced client-facing content; under the Reg BI Care Obligation at §240.15l-1(a)(2)(ii) the registered person must document the basis of every recommendation. The prompt sits in all three retention streams.

This has two operational consequences. First, prompts should be written as if a 2026 SEC examiner will read them next quarter — because they may. Second, prompts should be versioned the way the firm versions its IPS templates and its Form ADV Part 2A: a prompt has a name, a version number, an author, a creation date, a last-revised date, a tested-with-clients-N flag, and a regulatory regime tag (Marketing Rule, Reg BI, FINRA Rule 2210, internal-use-only). The Cardinal Rule from L1 Ch2.3 — source-system, regulatory, client-fit verification on every AI-touched artifact — extends backwards to the prompt itself: the prompt is the first artifact in the chain, and a sloppy prompt produces unverifiable downstream output.

The Five-Part Anatomy: Role + Context + Task + Format + Constraints

Every advisor prompt that produces defensible output has five parts. Drop any part and the output drifts. Drop two and the output hallucinates. Drop three and the output is so generic the advisor would have been faster writing it by hand.

Role

Role is the persona the model is asked to inhabit. Not "you are a helpful assistant" — that ships with the model and provides nothing. Instead: "You are a senior CFP-certificant advisor at a $750M RIA serving pre-retiree households with $1.5M-$5M of investable assets. You write in plain English, never invent facts, and cite every regulatory rule and IRC section you reference." The role constrains tone, vocabulary, citation habits, and risk posture. The L3 Ch9 lesson on Persona Engineering develops four locked personas (junior advisor, senior advisor, compliance reviewer, CCO) the firm uses repeatedly; for L2 Ch1 the discipline is to write the role explicitly every time, not to assume the model knows.

Context

Context is the household-specific or task-specific facts the model needs in order to produce non-generic output. Generic prompts produce generic answers. The Holistiplan-to-one-pager rewrite that produces a useful artifact for the Hendersons requires: "The Hendersons are a married couple, ages 64 and 62. 2024 MAGI was $148,000. Total IRA balances $1.6M (his $1.2M, hers $0.4M). Roth balances $190,000. Joint brokerage $610,000 (cost basis $410,000). Two grandchildren in a $90,000 529. RightCapital Monte Carlo 91% at current allocation; SS planning assumption is FRA on both spouses. IPS last updated 2022; revocable trust funding flagged 'open' since." With those facts, the model can identify the Roth conversion window, the IRMAA two-year-lookback bracket cliff, the SS-delay opportunity, and the unfunded-trust gap. Without them, the model can only produce platitudes.

Task

Task is the specific deliverable. "Summarize this Holistiplan output" is too vague. "Convert the attached Holistiplan tax-return summary into a one-page client-facing brief that names the three highest-leverage planning conversations for the 9:30 review, written in plain English at an eighth-grade reading level, that the clients can take home" is a task. Tasks should name the artifact, the audience, the reading level, and the length budget. Vague tasks are the single biggest source of unusable advisor output.

Format

Format is the shape of the output. "Three sections: (1) what changed since last year in two sentences, (2) the three planning conversations as a numbered list with one sentence each, (3) what we will do today as a numbered list with one sentence each. No headers larger than h3. No bullets inside the numbered lists. No marketing language. Total under 350 words." Format constraints make the output loadable into Wealthbox custom fields, the Holistiplan client-deliverable section, the RightCapital client portal, or the Pulse360 follow-up template. Structured output is the difference between "advisor uses AI" and "practice uses AI" — the L2 Ch8 lesson on structured output develops the JSON/Markdown discipline that loads cleanly into CRMs.

Constraints

Constraints are the negative space — what the model must not do. "Do not invent any number not provided in the context. Do not recommend any specific security or product. Do not cite any regulation by section number unless I have provided the section in context. Do not use marketing language ('powerful,' 'comprehensive,' 'proprietary,' 'best-in-class'). Do not promise outcomes. If a calculation requires data not provided, write '[need: X]' instead of inventing the number." Constraints are where the Marketing Rule and Reg BI live. The "do not invent" constraint protects against the SECURE 2.0 RMD-age hallucination and the wrong-2026-SS-maximum hallucination L2 Ch1 L2 catalogs. The "do not use marketing language" constraint protects against the AI-washing claims the 2024-2025 Delphia and Global Predictions settlements turned on. The "do not promise outcomes" constraint protects against the Marketing Rule's prohibition on misleading performance claims under Rule 206(4)-1(d).

Beginner Prompt vs. Senior Advisor Prompt on the Same Holistiplan Input

The pedagogy that makes the anatomy stick is the side-by-side. Same Holistiplan output, same household, two prompts.

The Beginner Prompt

"Summarize this tax return for my client meeting tomorrow."

The output is six paragraphs of competent-sounding tax-return narration. It explains what a 1040 line 11 is. It mentions that the client should "consider Roth conversions in years with lower income." It uses the phrase "tax-efficient retirement planning" twice. It does not name the bracket cliff. It does not name the IRMAA two-year lookback. It does not flag the unfunded trust. It does not produce a number. The advisor reads it, recognizes that nothing in it is wrong but nothing in it is usable, deletes it, and writes the brief by hand. Eleven seconds saved, twenty minutes wasted on a tab the AI never closed.

The Senior Advisor Prompt

"You are a senior CFP-certificant advisor at a $750M RIA. The Hendersons are a married couple, 64 and 62. 2024 MAGI $148,000. Total IRA $1.6M (his $1.2M, hers $0.4M). Roth $190,000. Joint brokerage $610,000 (cost basis $410,000). Two grandchildren in a $90,000 529. RightCapital Monte Carlo 91%. SS planning assumption FRA both spouses. IPS last updated 2022. Revocable trust funding flagged open since 2022. Attached: 2024 Holistiplan output.

Produce a one-page client-facing brief for the 9:30 review meeting today, structured as (1) Two sentences on what changed since last year. (2) Numbered list of the three highest-leverage planning conversations for today: name the conversation, state the dollar magnitude, name the deadline if any. (3) Numbered list of what we will do today: name the action, name the document, name the owner (me, client, custodian).

Constraints: Plain English at eighth-grade reading level. Under 350 words. Do not invent any number not in the context. Do not cite any IRC section or FINRA rule unless I have provided it. Do not use marketing language. Do not recommend any specific security or product. If a calculation requires data not provided, write [need: X] instead of inventing. Do not promise outcomes."

The output is a one-page brief naming (1) the Roth conversion window to the top of the 24% bracket (~$96,000 of conversion), (2) the IRMAA-bracket conversation about staying below $206,000 MAGI to avoid the next tier, (3) the SS-delay-to-70 conversation on the higher earner, the unfunded-trust action with the attorney handoff, and the IPS update task. Every number ties to the context. The advisor reads it, edits two sentences for tone, prints it, and walks into the meeting two minutes ahead of schedule. The difference between the two prompts is not the model — it is the anatomy.

The Draft-Critique-Regenerate Loop

No prompt is right the first time, even with the full anatomy. The discipline that turns a one-off prompt into a library artifact is the draft-critique-regenerate loop. The loop has four passes.

Pass 1: Draft on a Real Sample Client

Write the prompt against a real, anonymized client file. Not a hypothetical. Real prompts fail on the specifics: the spouse with $0 in IRAs (so the prompt's pro-rata-rule check fires when it shouldn't), the household with a defined-benefit pension (so the Roth-conversion math is wrong without including the pension income), the client with an inherited IRA on the 10-year SECURE 2.0 clock (so the RMD-calendar prompt produces wrong output unless the EDB classification is in the context). Run the prompt. Read the output. Find the first failure.

Pass 2: Critique Against the Cardinal Rule

Apply the L1 Ch2.3 three-tier verification to the output. Source-system: every number in the output should tie to the Holistiplan extract, the RightCapital plan, the Wealthbox household record, or the custodian feed. Regulatory: every rule cited should be cited correctly to the right regime (Reg BI §240.15l-1, Marketing Rule 206(4)-1, the January 2026 SEC staff FAQs, FINRA Rules 2210/3110/4511, the SECURE 2.0 RMD-age-73 vs. the deprecated 70.5 and 72, the IRC §408(d)(2) IRA aggregation rule read with §72(e)(8) basis recovery and reported on Form 8606 for backdoor Roth, the IRC §408(d)(8) QCD age-70.5 threshold, the IRC §402(e)(4) NUA mechanics, the May 2024 Reg S-P amendments). Client-fit: the output should be consistent with the household's IPS, prior decisions, and stated preferences. Note every failure in writing.

Pass 3: Regenerate With Fixes

Edit the prompt to close each failure. If the model invented the 2026 Social Security maximum taxable earnings number, add a constraint: "Do not state any annual IRS, SS, or Medicare figure unless I have provided it in the context." If the model cited an old RMD age, add a constraint: "Under SECURE 2.0, the RMD age is 73 (and will rise to 75 in 2033). Do not use 70.5 or 72." If the model used marketing language, add to the role: "Do not use words from this list: powerful, comprehensive, proprietary, best-in-class, cutting-edge, AI-powered, machine-learning-driven, algorithmic." Run the prompt again. Read the output. Find the next failure.

Pass 4: Lock and Version Into the Library

When the prompt produces clean output on three to five consecutive real samples, lock it. The lock metadata: prompt name (e.g., "Annual-Review-Brief-v2.3"), version number, author, creation date, last-revised date, tested-with-clients-N flag (e.g., "tested with 7 households, May 2026"), regulatory regime tag (Reg BI internal-use, not a marketing communication), source artifact types (Holistiplan output, RightCapital plan, Wealthbox household, Orion portfolio), and the named-tool downstream destinations (Wealthbox custom fields, Holistiplan client deliverable, Pulse360 follow-up). Store the prompt in the firm's RAG-connected document vault (L3 Ch9 develops the retrieval-augmented workflow) or — for a smaller practice — in a versioned shared file with a changelog. The L2 capstone deliverable is a 25-prompt library; every prompt in it carries the lock metadata.

The Hendersons End-to-End: One Prompt's Evolution Across the Loop

To make the loop concrete, here is one prompt's evolution across four passes on the Henderson household.

v1.0 (beginner). "Summarize this 1040 for my meeting." Output: generic paragraph. Failures: no client-specific facts, no number magnitude, no Roth window, no IRMAA, no trust gap.

v2.0 (anatomy applied). The senior prompt above. Output: client-facing brief with the three conversations. Failures on first sample: the model invented "the Hendersons should consider a 72(t) substantially equal periodic payment to bridge to Social Security" — neither client is under 59.5 nor needs penalty-free access. The 72(t) recommendation does not fit and is unsolicited.

v2.1. Added constraint: "Do not recommend strategies that do not address a stated household need. If you propose a strategy, state in one phrase the household-specific need it addresses." Re-run: model dropped the 72(t) and added the SS-delay conversation correctly. Failure on next sample (a 71-year-old client): the model proposed a Roth conversion to the top of the 24% bracket without checking whether the client was already past the IRMAA cliff that year, which would have meant the conversion cost two IRMAA tiers.

v2.2. Added constraint: "Before recommending a Roth conversion, check (a) current-year MAGI vs. IRMAA bracket thresholds two years forward, (b) whether the client is currently in a Roth contribution window or a conversion window only, (c) whether pre-tax IRA balances will trigger the pro-rata rule under IRC §408(d)(2) read with §72(e)(8), tracked on Form 8606, on any backdoor contribution. State each check explicitly in the brief." Re-run on five households: clean. Locked as Annual-Review-Brief-v2.3, tested-with-clients-5, May 2026, internal-use only, source artifacts (Holistiplan, RightCapital, Wealthbox), downstream (Wealthbox activity log, printed client brief).

The locked prompt now runs across the book. The advisor's marginal time on a review brief drops from twenty minutes to forty-five seconds. The prompt is a Reg BI artifact retained under Rule 4511, version-stamped, with a documented test history. The next FINRA examiner who asks "how do you produce your client review briefs?" receives the prompt, the version log, the test history, and the verification protocol. The answer is short, accurate, and survives.

Building the Firm's Prompt Library — Naming, Versioning, and Governance

One advisor's prompt library is a productivity asset. A practice's prompt library is competitive infrastructure. The L2 capstone delivers 25 prompts; the L3 Ch1 workflow audit identifies the next 25; the L4 Ch3 WSPs govern who can change them and how. The naming convention that survives a firm of 8-30 advisors: {Workflow}-{Artifact}-v{Major}.{Minor}. Workflow examples: Annual-Review, Discovery, Rollover-RegBI, Roth-Conversion-Memo, RMD-Calendar, Estate-Gap-Audit, NUA-Decision, QCD-Workflow, IPS-Draft. Artifact examples: Brief, Memo, Email, Checklist, Talking-Points. Version increments: major version for change of structure or audience, minor version for constraint additions or critique fixes.

Governance: every locked prompt has an owner (the lead advisor or CCO who can approve changes), a reviewer (a second advisor or compliance lead who must sign off on changes that touch client-facing language or regulatory citations), and a usage log (a quarterly review of which prompts are actually used and which are dead weight). The L3 Ch10 lesson on the Zocks-to-Wealthbox-to-Smarsh pipeline integrates the prompt library into the archive: every prompt invocation produces a retained record (prompt, output, edits, signoff) under SEC Rule 204-2 and FINRA Rule 4511. The L4 Ch7 lesson on Marketing Rule strategy treats the prompt library as part of the substantiation file — when a regulator asks "how do you produce your AI-generated client content?" the firm answers with the library, the lock metadata, and the principal-review queue.

Putting the Anatomy to Work Monday Morning

The first deployment of this lesson is small. Pick one recurring workflow — annual review brief, follow-up email, IPS section draft — and write a single prompt with the five-part anatomy. Run it on three real anonymized client files. Apply the Cardinal Rule's three-tier verification to the output of each. Iterate the prompt twice. Lock it with the metadata and store it in a shared file with a changelog. That one prompt, locked and used twenty times a quarter, saves more time over a year than every "AI 101" webinar combined and produces output that survives the Reg BI exam the L4 capstone teaches you to host. The next lesson — Recognizing Bad Output Before It Hits a Client — installs the failure gallery that makes the critique pass faster.

Key Takeaways

• Every advisor prompt has five parts — Role, Context, Task, Format, Constraints. Drop one and output drifts; drop two and it hallucinates; drop three and it's faster to write the artifact by hand.
• Prompts are Reg BI artifacts retained under SEC Rule 204-2 and FINRA Rule 4511. Write them as if a 2026 examiner will read them next quarter, because they may.
• The beginner-vs-senior prompt contrast on the Hendersons shows that the model is not the bottleneck — the prompt is. Same model, same Holistiplan input, two completely different briefs.
• The draft-critique-regenerate loop has four passes — draft on a real sample, critique against the Cardinal Rule's three-tier verification, regenerate with fixes, lock and version into the library with name, version, author, dates, tested-with-N flag, and regulatory regime tag.
• Constraints are where the Marketing Rule and Reg BI live — "do not invent" closes hallucination, "do not use marketing language" closes AI-washing under 206(4)-1, "do not promise outcomes" closes performance misrepresentation under 206(4)-1(d).
• A firm prompt library is competitive infrastructure, not a personal productivity asset. Naming convention {Workflow}-{Artifact}-v{Major}.{Minor}, owner + reviewer, quarterly usage review, integrated into the Smarsh archive pipeline of L3 Ch10 and the substantiation file of L4 Ch7.
• Monday deployment: one prompt, three real samples, two iteration cycles, locked with metadata. That single locked prompt, used twenty times a quarter, beats every AI 101 course you could attend this year.