Variance narratives — the "why" behind the numbers

TL;DR. Compare actuals to budget, identify the material movers, and write the narrative paragraphs that go to leadership — in the controller's voice, not Cowork's. Around 3 hours per close saved at steady state. The hardest part is the voice; train it with three good and three bad examples in CLAUDE.md.

Job to be done#

Compare actuals to budget or last period, identify material variances, and write the narrative paragraphs that accompany the close pack.

Who runs it#

FP&A lead or finance business partner.

Inputs (inbox/)#

• Actuals vs budget Excel (P&L by department or product)
• Driver definitions: which lines need full-paragraph commentary, which need a sentence
• House-style examples from prior months

Outputs (output/)#

• variance-commentary.docx with one paragraph per material driver
• Underlying decomposition Excel (price × volume × mix)
• A stoplight summary table — red / amber / green by line

Prompt seed#

Read /inbox/actuals-vs-budget-[month].xlsx.
For each line item with absolute variance > £25k OR > 5%,
write one paragraph explaining the driver, in the voice and structure
of /CLAUDE.md/style-examples.docx.
Include the magnitude (£ and %) and a one-sentence forward look.
Output to /output/variance-commentary-[month].docx and an underlying
decomposition in /output/variance-decomp-[month].xlsx.

Quality bar#

• Every paragraph cites the magnitude exactly; no rounding.
• Cause language is appropriately hedged ("driven by", "primarily due to"). Cowork tends to over-claim causality. Push back.
• House style — no marketing language, no business-school jargon.
• Common trip-up: drivers Cowork cannot actually know ("delayed by supplier issue"). Leave a [CONFIRM] placeholder rather than inventing the cause.

Time saved (typical)#

About 3 hours per close at steady state.

Upgrade path#

• Convert to a variance-narrative skill referencing the saved style examples.
• Add a finance-warehouse connector for live data.

Tinkso's take#

The hardest part is the voice. Cowork's default tone is too analytical and too hedged for most CFO audiences. Train it with three good and three bad examples in CLAUDE.md and the voice locks in by month two. Without those examples, expect six weeks of "can you make it sound less like a robot" feedback loops.