A prompt is a work order, not a wish
Last session you learned the universal method: ground it, make it show its work, verify. This session is where you get good at the "ground it" half. Most disappointing AI output in finance is not a model problem - it is a briefing problem. You would never ask a junior analyst to "comment on these results" with no numbers, no format, and no deadline; a chat AI deserves the same respect. Tonight you write prompts the way you would brief a sharp new hire: role, context, data, task, format, and the rules they must not break.
Anatomy of a reliable finance prompt 6 min live
A strong finance prompt has six layers, and they stack. Drop one and the answer gets vaguer, more invented, or wrongly formatted. You do not need all six every time - but knowing the full stack means you know exactly which layer to add when an answer disappoints.
LiveThe six layers, in the order you write them3 min▶
Write them top to bottom and the prompt almost drafts itself. Each layer answers one question the model would otherwise guess at.
- Role - "You are a controller / FP&A analyst / credit analyst." Sets vocabulary and rigour. A model told it is an analyst hedges and shows work; told nothing, it chats.
- Context - the business, the period, the audience for the answer. "This is for a board pack" produces different prose than "explain to me".
- Data - paste the actual numbers. This is rail-one grounding from b1. Public or synthetic data in a consumer tool; real financials in enterprise or local only.
- Task - one clear verb: explain, compare, reconcile, draft. Two tasks in one prompt gives you two half-answers.
- Format - table, bullets, JSON, word count. Say it up front and you stop re-asking "now put that in a table".
- Constraints - the honesty rules: show every calculation, use only the numbers supplied, say "not in the data" instead of inventing, and state any assumption made.
Self-studyThe constraint lines, ready to paste2 min read▶
Keep these four lines on a sticky note. They are the difference between an answer you trust and one you re-check line by line.
Paste this at the end of any analytical prompt. It costs you four lines and buys you a checkable answer. Everything after this in the track assumes you are using something like it.
The core patterns - five moves that cover most work 6 min live
You do not need fifty prompt tricks. Five patterns handle the overwhelming majority of finance asks, and they combine. Learn when each one earns its place and you stop guessing.
LivePattern picker - when to reach for each3 min▶
| Pattern | What it does | Reach for it when... |
|---|---|---|
| Few-shot (show an example) | Give one sample comment in your house style; the model matches it | You want consistent tone and structure across many outputs |
| Role prompting | "You are a controller reviewing the close" | You want the rigour and vocabulary of a specific finance role |
| Step-by-step decomposition | "First compute variances, then rank, then comment on the top 3" | The task has stages and you want the math visible and ordered |
| Output control | "Return a 3-column table" or "return valid JSON with these keys" | The answer feeds a doc, a sheet, or another system |
| Self-check | "List your assumptions and what I should verify before I use this" | Always - it is your built-in rail-two review step |
Few-shot is the quiet hero. One good example of a variance comment teaches the model your style faster than three paragraphs of instructions ever will.
Self-studyFew-shot and self-check, shown2 min read▶
Few-shot means you supply the pattern, not just the instruction. Watch how one example locks the style:
Self-check makes the model hand you its own review notes - the single cheapest way to honour rail two inside the prompt itself:
Finance guardrails inside the prompt 3 min live
General prompting advice ignores the things that get finance people in trouble: the wrong currency, a quarter compared to a full year, or a percentage rounded three different ways in one memo. Bake the hygiene into the prompt so the model cannot drift.
LiveFive lines that stop silent finance errors2 min▶
- Currency and units. "All figures in SGD thousands" - or the model may silently mix dollars and thousands, or assume USD.
- Period labels. Spell out Q1 vs FY, actual vs budget, this year vs last. A comparison against the wrong base is worse than no comparison.
- Rounding. "Round to whole numbers; percentages to one decimal." Consistency reads as competence in a board pack.
- Materiality. "Only comment on variances above 50k or 10%." Stops the model narrating trivia and burying the real story.
- Only these numbers. The b1 grounding rule, restated every time: no outside figures, no memory, no filling gaps.
The memo that mixed thousands and millions. An analyst pasted a P&L in thousands but asked for commentary "in millions" - the model gamely reported revenue of "4.1 million" when it was 4.1 thousand-times-something-else. It looked authoritative and was off by orders of magnitude. One line - "all figures in SGD thousands, do not convert" - would have prevented it. Guardrails are cheaper than corrections.
Rebuild "comment on these results" into a real prompt ★ 12 min · everyone builds
The most common finance prompt in the wild is also the worst: "comment on these results". We take that lazy ask, run it, feel the vagueness, then rebuild it layer by layer on a small supplied P&L - and iterate once to sharpen it.
Open ChatGPT or Claude. Paste the small P&L below (it is synthetic - safe for any tool, rail one). You will use it for both prompts.
Run the vague version first (Prompt A). Notice what you get: generic prose, maybe invented drivers, no math shown, no idea which variances actually matter.
Now rebuild it with the six layers (Prompt B): role, context, the data, one clear task, a table format, and the constraint block. Run it.
Compare. The structured answer ranks variances by size, shows the subtraction, respects your materiality threshold, and flags what to verify. That is the same fiction-to-checkable jump from b1, now on purpose.
Iterate once. Add a single refinement - "add a one-line 'so what' for the CFO under the table" - and re-run. Feel how cheap it is to steer once the structure is right. That is the whole skill.
Bottle it: a reusable variance-commentary template ★ 8 min · build your own
A great prompt you have to rewrite every month is a chore. A great prompt with placeholders is a tool. We turn Prompt B into a template you fill in seconds - the first entry in your prompt library.
Take the structure of Prompt B and replace the specifics with clearly marked placeholders in square brackets: role, period, currency, the data table, and the materiality threshold.
Save it somewhere you will find it - a notes doc, a shared team page, or your tool's saved-prompts feature. Give it a name: "Variance commentary v1".
Test the template by filling it with this month's numbers. If a placeholder is unclear, tighten the wording now, while you remember what you meant.
Start a running "prompt library" file. Every reusable prompt from this track goes in it - by b10 you will have a small, battle-tested toolkit no one handed you.
From two hours to twenty minutes. A group FP&A analyst templated her monthly variance pack this way. The first draft used to eat a morning; now she pastes the numbers, runs the template, and spends her time on the judgement - checking the math and sharpening the "so what". The AI did not replace her. It deleted the boring 80% so she could do the 20% that is actually her job.
Try it yourself - this week ◐ 30-45 min total
- Finish both demos if you ran short in class - especially the rebuild. Feeling the jump from Prompt A to Prompt B is the lesson.
- Take one prompt you actually use at work and add the missing anatomy layers. Note which layer was missing - that is usually your habitual blind spot.
- Build one more template from your real recurring work: a flash-report prompt, a reconciliation-explainer, or a KPI-summary. Add it to your prompt library.
- Try the few-shot pattern once: give the model one example of your house style and see how much closer it lands than instructions alone.
- Optional: swap templates with a colleague. The fastest way to improve a prompt is to watch someone else fill it in and get confused.
Official sources covered
This session teaches from public prompting curricula for finance. Certificates, graded assessments, and paid tooling stay with the official programs - noted honestly. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · What are the six layers of a reliable finance prompt?
Role sets the standard, context fits your world, data grounds it, task is the one job, format shapes the output, and constraints carry the honesty rules. Drop one and the answer drifts.
2 · You need the same style of variance comment across twenty line items. Which pattern helps most?
One concrete example teaches style faster than any amount of instruction. Few-shot is the quiet workhorse of consistent finance output.
3 · Why add a self-check line ("list your assumptions and what I should verify")?
Self-check hands you the model's own soft spots to verify. It does not replace your review - it focuses it. That is rail two, baked into the prompt.