Module 4 — The Workflow Prompt Library, Part One

How to use this library

The next two modules are the working core of the course. Each workflow follows the same four-part structure so you can drop it straight into practice: what the workflow is and where AI genuinely helps, a copy-paste-ready prompt built on the Module 3 structure, what good output looks like so you can recognize it, and the review-and-guardrail step that keeps the professional in the loop. Every prompt assumes you have de-identified the material per Module 2 and that you will review the output before it is used.

One recurring note runs through both modules: model choice by task. Frontier models — the largest, most capable, most expensive — earn their cost on tasks with high ambiguity, long context, real reasoning, or client-facing stakes. Cheaper or smaller models, including local models that run on your own hardware and never send data anywhere, are entirely sufficient for well-structured, low-ambiguity transformation: categorizing transactions against a fixed list, reformatting, simple extraction from clean input. We flag which is which in each workflow, because matching the model to the task is how you get quality where it matters and keep cost trivial everywhere else. Module 6 develops the economics; here we just mark it.

Workflow 1 — Engagement letters and scope drafting

The workflow and where AI helps. Engagement letters are high-value, repetitive, and structurally similar across clients, which makes them ideal for AI-assisted drafting. You are not asking the model to invent your firm's terms; you are asking it to assemble a first draft from your standard components, tailored to a described engagement, so you spend your time reviewing and adjusting rather than starting from a blank page. AI helps most with the tailoring — adjusting scope language to a particular service, spelling out deliverables and exclusions, and producing a clean, readable draft in your firm's structure.

The prompt.

You are drafting a professional services engagement letter for an accounting
firm. Use the firm's standard structure below and tailor it to this
engagement. Do not invent fee amounts, legal terms, or regulatory language;
where a specific term or number is needed, insert a clearly marked
[PLACEHOLDER] for the firm to complete.

Firm standard sections: Introduction and parties; Scope of services;
Client responsibilities; Deliverables and timeline; Fees and billing;
Limitations and exclusions; Term and termination; Signatures.

Engagement details:
- Service: [e.g., monthly bookkeeping and quarterly financial statement
  preparation]
- Client type: [e.g., single-location retail LLC]
- Period covered: [e.g., calendar year 2026]
- Notable inclusions: [list]
- Notable exclusions: [e.g., tax return preparation is NOT included]
- Any technology-use disclosure the firm wants stated: [paste firm's
  standard AI/technology-use clause here, if any]

Produce the draft with clear section headings. In the Limitations and
Scope sections, be explicit about what is excluded. Flag with [REVIEW] any
sentence where you were uncertain about the firm's intent.

What good output looks like. A cleanly sectioned draft that reads in your firm's register, with scope and exclusions stated plainly, every fee and legally operative term left as a marked placeholder rather than invented, and [REVIEW] flags wherever the model guessed at intent. Notice we explicitly instruct it not to fabricate legal or fee language — those are exactly the specifics a model will otherwise confidently invent.

Review and guardrails. An engagement letter is a contract; it is reviewed and finalized by a qualified person, and where terms have legal weight, by or with counsel per firm policy. Confirm every placeholder is filled, every exclusion is correct, and that the technology-use disclosure (Module 2) is present if your firm includes one. The AI produced a draft; the firm owns the agreement. A frontier model is worth using here — the language is client-facing and the tailoring benefits from stronger drafting — but the review is where the document becomes real.

Workflow 2 — PBC and document-request list generation

The workflow and where AI helps. The "prepared by client" list — the request for the documents and information you need to begin an engagement — is tedious to build from scratch and easy to leave incomplete, and an incomplete request means a mid-engagement scramble. AI is well suited to generating a thorough first-pass request tailored to the engagement type, because it can draw on the general shape of what such work requires and organize it into a clean, categorized list you then prune and adjust to the specific client.

The prompt.

You are preparing a document-and-information request list ("PBC list") for
the following engagement. Produce a categorized, checkbox-style list of the
documents and information the firm will typically need from the client to
perform this work. Organize by category. For each item, add a brief note on
why it is needed. Mark any item that is only conditionally needed with
"(if applicable)".

Engagement: [e.g., preparation of annual financial statements for a
construction subcontractor, calendar year 2026]
Known specifics: [e.g., cash basis; one bank account; ~15 employees;
owns equipment and one vehicle]

At the end, add a section titled "Consider asking about" listing 5-8 items
that are commonly overlooked for this type of engagement but may not apply.
Do not assume facts about this client beyond what I provided.

What good output looks like. A categorized checklist — bank and financial records, payroll, fixed assets, loans, prior-period workpapers, and so on — each item with a one-line rationale, conditional items flagged, and a closing "consider asking about" section that catches the things people forget. It reads like a complete starting request that a professional trims rather than expands.

Review and guardrails. Prune it. The model errs toward over-requesting, and asking a client for things you do not need wastes their time and yours. Confirm the list matches the actual scope and add anything client-specific the model could not know. This is a strong candidate for a cheaper model — the task is well-structured and low-ambiguity, and a mid-tier or even local model produces a solid list. The human step is editorial trimming, not error-hunting, which is a comfortable place to be.

Workflow 3 — Workpaper review and tie-out sanity checks

The workflow and where AI helps. AI can serve as a fast second set of eyes on a workpaper's internal consistency — does the total foot, do the subtotals sum to the total, do figures that should agree across schedules actually agree, are there obvious transpositions or sign errors. It does not replace your review and it is not a calculator you can trust for the arithmetic itself, but it is good at spotting where to look and at articulating what a discrepancy might be. Use it to triage, then verify every flag yourself.

The prompt.

You are reviewing an accounting workpaper for internal consistency. Below is
the data. Do the following checks and report only what you find:

1. Check whether the line items sum to the stated total. State the sum you
   calculate and the stated total, and the difference if any.
2. Identify any subtotal that does not equal the sum of its components.
3. Flag any figure that appears in two places but with different values.
4. Flag likely transposition errors (e.g., 5,400 vs 4,500) or sign errors.

Show your arithmetic for every check so I can verify it. If everything ties,
say so explicitly. Do not "correct" anything silently — report discrepancies
for me to resolve.

[paste the de-identified workpaper data]

What good output looks like. A short report that states the computed sum against the stated total with the difference, names any subtotal mismatch, points to any cross-reference disagreement, and shows the arithmetic for each so you can check it. Where everything ties, it says so rather than inventing a problem to seem useful.

Review and guardrails. This is the workflow where the calculator caution from Module 1 bites hardest: do not trust the model's arithmetic — verify every number it computes. Its value is direction ("look at this subtotal"), not computation. Treat a "no discrepancies" result as a prompt to confirm, not as clearance. And because the model can miss a real error while confidently reporting none, this is a triage aid layered on top of your normal tie-out, never a substitute for it. A frontier model reasons better about what kind of error a discrepancy suggests, but even it must have its arithmetic checked.

Workflow 4 — Bank and GL reconciliation triage and exception explanation

The workflow and where AI helps. Reconciliations generate exceptions — items on one side without a match on the other — and much of the work is explaining, per exception, the plausible reason: timing, a transposition, a missing entry, a duplicate. AI is good at proposing candidate explanations for each exception and grouping them, turning a raw list of unmatched items into a triaged list with hypotheses to check. The matching itself is often better handled by your reconciliation software; AI shines at the narrative triage layer on top.

The prompt.

You are helping triage reconciliation exceptions between a bank statement and
a general ledger cash account. Below are the unmatched items from each side.
For each unmatched item, propose the most likely explanation(s) from these
common causes: timing difference (deposit/payment in transit), transposition
error, missing journal entry, duplicate entry, bank fee not recorded,
or other (specify). Where two items across the sides might be the same
transaction, pair them and explain why.

Output a table: Item, Side (Bank/GL), Amount, Likely cause, Suggested action
to verify. Rank the whole list by which exceptions are most likely to be
real errors versus timing. Do not conclude a cause is certain — these are
hypotheses for me to confirm.

[paste de-identified unmatched items from both sides]

What good output looks like. A ranked table pairing likely-matching items across the two sides, a plausible cause and a concrete verification step for each, and errors sorted ahead of timing differences so you work the real problems first. Every cause is framed as a hypothesis, not a finding.

Review and guardrails. Each proposed explanation is a lead you confirm against the actual records; a plausible cause is not a resolved item. Watch for the model pairing two items that are not actually the same transaction just because the amounts are close — coincidental matches are exactly the trap. The reconciliation is not done until you have verified each item; the AI has organized the work, not completed it. A mid-tier model handles this well when the exception list is clean; reach for a frontier model when the transactions are complex or the descriptions are cryptic.

Workflow 5 — Client email, memo, and advisory drafting

The workflow and where AI helps. Much of an accountant's written communication is explaining something technical to a non-technical client in a clear, appropriately warm, professional register — and this is squarely in an LLM's wheelhouse, because it is pure language transformation from your notes into polished prose. AI gives you a strong first draft of a client email or memo from a few bullet points of substance, which you then correct and finalize. The substance must come from you; the model supplies the drafting speed.

The prompt.

You are drafting a client-facing email on behalf of an accountant. Write in
a professional, clear, and warm tone suitable for a small-business owner who
is not an accountant. Explain the substance below in plain language, avoiding
jargon or defining it briefly where unavoidable. Keep it concise.

Recipient context: [e.g., owner of a landscaping business, not financially
sophisticated, prefers directness]
Substance to convey (these are the facts — do not add to them):
- [bullet 1]
- [bullet 2]
- [any action you need the client to take, with a deadline if relevant]

Produce: a subject line and the email body. Do not invent figures, dates, or
commitments beyond what I listed. End with a clear next step if one is needed.

What good output looks like. A ready-to-edit email with a useful subject line, a plain-language explanation that stays strictly within the facts you supplied, jargon either avoided or briefly defined, and a clear call to action. Because you instructed it not to add facts, it should not have invented a figure or a deadline you did not give it.

Review and guardrails. Read it as though a junior drafted it, because one effectively did. Confirm every factual statement, since the model can still slip in a plausible detail you did not provide; confirm the tone fits this particular client relationship; and make sure nothing commits the firm to something you did not intend. For advisory content specifically, ensure the draft does not overstate certainty or stray into advice the facts do not support — you, not the model, own the professional judgment in the message. A frontier model produces noticeably better client-facing prose, and the modest cost is justified when the reader is a client. As always, minimize what you paste: the client's name and specifics can be added by you after drafting rather than sent to the model.

A shared review discipline for every workflow

Before moving on, it is worth naming the review habits that apply across all ten workflows, because they are the same few moves in each and stating them once lets us assume them everywhere. Whatever the task, a professional review of AI output does four things.

It checks the output against the source. Every workflow here transforms material you provided, so the first question is always whether the output faithfully reflects that material — whether a summarized figure actually appears in the document, whether a categorized item was categorized per the real rule, whether a drafted email stayed within the facts you supplied. This is the check that catches hallucination, and it is fast precisely because the source is right there. An output you cannot check against a source is an output you should not have asked for.

It checks the arithmetic separately. Any number the model computed is unverified until you confirm it, because the model is not a calculator. This is a distinct step from checking against the source, and skipping it is how a plausible wrong total reaches a workpaper.

It checks for confident omissions. The dangerous errors are often not what the model got wrong but what it silently left out — the exception it did not flag, the asset it skipped, the caveat it dropped. Reviewing for completeness, not just correctness, is what catches these, and the guardrail phrasing from Module 3 ("list anything you could not determine") makes them visible.

It applies professional judgment the model cannot have. The model does not know your client, your firm's positions, or the current authority. The final layer of review is bringing that knowledge to bear — asking whether the output makes sense for this client, in this situation, under the rules as they actually stand. This is the layer that is irreducibly the professional's, and it is why the review is the profession rather than an inconvenience.

Run those four moves and the review is real. Skip them and you have a rubber stamp, which is worse than no AI at all because it launders a guess into a signed deliverable. Every workflow in this module and the next assumes all four.

A word on where these fit in an engagement

None of these workflows is a product you hand a client; each is an internal accelerant that produces a better or faster starting point for work you were going to do anyway. The engagement-letter draft still becomes a firm-owned contract; the reconciliation triage still feeds your normal reconciliation; the client email still goes out under your name after your edit. Keeping that framing clear prevents the most common adoption mistake — treating an AI first pass as a finished work product because it looks finished. It looks finished because fluency is what the model is best at; whether it is finished is what your review determines. The time these workflows save is real, and it is saved at the front of the work, where a blank page or a tedious sort used to sit. The judgment, the accuracy, and the responsibility stay exactly where they were.

Carrying forward

These five workflows share a shape you will see again in Module 5: the AI transforms material you supply into a structured, usable first pass, and a professional reviews it into a deliverable. Where the task is well-structured and low-stakes — the PBC list, simple categorization — a cheap or local model suffices and the human step is light editing. Where the task is client-facing or reasoning-heavy — engagement letters, advisory drafts, complex exception triage — a frontier model earns its cost, and the human step is substantive review. In every case the guardrail is the same one Module 2 made non-negotiable: the model assists, the professional is responsible. Module 5 continues with five more workflows, including the cost segregation anchor worked in full and the citation-discipline workflow that is the profession's most important defense against fabricated authority.