SuperAPI AI Directory

variance-commentary

General & other

View repository

Updated 10-15 20:11

About this skill

Variance commentary

Given current-period actuals, prior-period actuals, and budget for the same scope, produce a commentary table.

Threshold

Flag a line for commentary if either is true:

  • Absolute variance ≥ the firm's materiality threshold (use the provided value; default 5% of the line or a fixed floor, whichever is greater)
  • The line is on the "always comment" list (revenue, headcount cost, cash)

For each flagged line

Column Content
Line Account or caption
Current / Prior / Budget The three values
Δ vs prior and Δ vs budget Amount and %
Driver One sentence explaining the movement from underlying activity — not a restatement of the number

A driver explains why, not what: "Cloud spend up $1.2M on incremental GPU reservations for the May launch" — not "Cloud spend increased $1.2M (18%)."

Sourcing the driver

Look at the activity behind the line (journal-source breakdown, vendor mix, headcount delta, volume × rate) via the internal-gl MCP. If the driver isn't clear from the data, write "driver unclear — flag for controller" rather than inventing one.

Output

The commentary table plus a short narrative (3–5 sentences) summarizing the period's biggest movers.

Hermes Profile Notes

This skill was packaged from pi-financial-services source path plugins/vertical-plugins/fund-admin/skills/variance-commentary for the Hermes financial-services profile. Use institutional data connectors first when available, cite sources, and stage outputs for qualified human review.

Common Pitfalls

  1. Do not present drafts as investment, legal, tax, or accounting advice.
  2. Do not use web search as the primary source when an institutional MCP/data connector is available.
  3. Do not execute transactions, contact clients, post to a ledger, or approve onboarding. Stage outputs for review.

Verification Checklist

  • Inputs, assumptions, dates, and currencies are explicit.
  • Numbers tie across tables, models, decks, and memos.
  • Sources are cited and institutional data is preferred where available.
  • Output is labeled draft / for human review where appropriate.