Board pack preparation is one of the most punishing recurring tasks in a finance function. The data arrives late, in inconsistent formats, from six different owners. The narrative has to reconcile to the numbers to the decimal. And the audience is unforgiving: directors who read the pack once, on a Sunday evening, and form their entire impression of management’s grip on the business from those pages. Most finance teams respond by throwing headcount at it — analysts rebuilding the same charts every month, controllers rewriting the same commentary, and a CFO doing a final quality pass at 11pm the night before the meeting.
The friction is not analytical. It is structural. Board packs are an assembly problem: dozens of source files, a fixed template, a house style nobody has ever written down, and a hard deadline. This is precisely the kind of work large language models handle well — not because they can do the thinking, but because they can absorb the structure, hold the rules, and draft against a reference faster than any human reviewer can brief a colleague.
Used properly, ChatGPT does not write your board pack. It compresses the distance between raw subsidiary submissions and a first draft that a CFO can actually edit. The key word is properly: an unstructured “summarise these results for the board” prompt produces generic, hedged prose that creates more rework than it saves. What follows is the structured approach — two prompt templates that treat the model as a briefing-driven drafting engine with explicit success criteria, reference patterns, and a mandatory clarification step before it writes a single word.
The two-stage workflow: narrative first, numbers second
Board packs fail in two distinct places, and they need two distinct prompts. The first failure is the narrative: commentary that restates the numbers instead of explaining them, buries the variance story, or drifts from the house tone. The second failure is the numbers layer itself — inconsistent definitions, unexplained bridges, tables that do not tie to the commentary. Stage one produces the MD&A and section commentary. Stage two stress-tests the quantitative layer against the narrative before anything goes to the CFO.
Both prompts use the same anatomy: a task statement with success criteria, an explicit file-reading instruction, a reference example with extracted patterns, a success brief, and a hard stop for clarifying questions before execution. The clarifying-questions step is not a formality. It is where the model surfaces the gaps in your own briefing — the missing prior-period comparator, the undefined “underlying” adjustment — before those gaps become errors in the pack.
First, read these files completely before responding:
[prior_board_pack_Q2.md] — last quarter’s full pack, including MD&A, segment commentary, and treasury update
[house_style_guide.md] — tone rules, terminology, banned phrases, table conventions, and how we refer to segments and entities
[current_period_data.md] — this month’s P&L, balance sheet, cash flow, and KPI table by segment, with prior month and budget comparators
[management_notes.md] — bullet notes from the CFO and segment leads on what drove the month
Here is a reference for what I want to achieve:
[Upload the MD&A section from prior_board_pack_Q2.md as the reference]
Here’s what makes this reference work:
Variance commentary leads with the driver, not the number. Every material variance is explained in one sentence, then quantified. Segment sections follow a fixed order: performance, drivers, outlook. Tone is direct and unhedged; no “we are pleased to report.” Tables carry a single footnote convention. Forward-looking statements are flagged explicitly.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: MD&A plus four segment commentaries, approximately 1,200 words total
Recipient’s reaction: the CFO should read it and think “this is 90% there” — edits, not rewrites
Does NOT sound like: a press release, a data dump, or generic AI commentary (“strong performance across the board”)
Success means: fewer than five substantive CFO edits and zero factual reconciliations flagged
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
Note what the first prompt does not ask for: creativity, insight, or strategic recommendations. It asks for pattern-matching against a reference the model has read in full, constrained by an explicit success measure — edit count. That measure is what keeps the output usable. Without it, you get prose that is pleasant to read and useless to a board.
First, read these files completely before responding:
[draft_board_pack.md] — the full draft pack for this quarter, all sections and appendices
[definitions_and_policies.md] — our official definitions for adjusted EBITDA, underlying growth, net debt, and segment allocation rules
[data_workbook_export.md] — the underlying numbers as exported from the consolidation system
Here is a reference for what I want to achieve:
[Upload last quarter’s post-meeting corrections log as markdown]
Here’s what makes this reference work:
Each correction is logged as: location, issue type (tie-out, definition, narrative conflict, rounding), what was wrong, and the corrected value. Issues are ordered by materiality. Nothing is flagged without a specific page or table reference and a proposed fix.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: a correction log table, one row per issue, plus a short summary of the three most material risks
Recipient’s reaction: the controller should be able to work the list top to bottom without asking a single clarifying question
Does NOT sound like: a general review, style commentary, or vague observations (“consider reviewing the cash flow section”)
Success means: every narrative figure traces to the workbook, every definition matches policy, and no issue is raised without a proposed correction
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
Run the second prompt after the narrative draft is merged into the pack, not before. It works because the model is asked to compare three artefacts against each other — draft, definitions, source data — rather than to exercise judgement about what the board wants. Consistency checking is mechanical, and mechanical is where models are strongest. The output is a worklist, not an opinion, which is exactly what a controller needs at that stage of the cycle.
Two practical tips before you try this. First, build the two context files — house style guide and definitions and policies — once, and maintain them. They are the highest-leverage documents in the whole workflow; every prompt run inherits their quality. Second, treat the clarifying-questions step as a free review of your own briefing: if the model asks something you cannot answer, the board will probably ask it too. Start with the narrative prompt this quarter, compare the draft against your last pack, and only add the consistency audit once the first stage is reliably producing single-edit-pass output.
Published on 18 October 2026 on growwithgpt.com