Use cases, tested before they're published.

We built a prompt that guards against five failure modes and tested it against a plain request. It didn't make the decks more accurate: on messy exports, 54 of 60 plain decks had every figure right against 49 of 60 with the prompt, at 1.6× the output tokens. What did move the result was the model, and checking the numbers.

1. 1 · Pick the model for the job

   Claude Opus 5.5 had every figure right on 40 of 40 messy decks. The cheaper models missed on some, mostly arithmetic and blank cells. For numbers leaders will act on, the model was the biggest lever in our tests.

2. 2 · Ask plainly

   Say the period, what counts (actuals, not forecast) and what to leave out. A plain request did as well as our prompt. One of the prompt's own rules caused errors: told never to infer a mapping, two decks refused a region the workbook documents as renamed.

3. 3 · Check the deck before it goes out
   Without code: five checks, ten minutes

   • Recompute the total. One SUM in Excel, with the same filters you asked for. Compare it to slide one. fmi_4_1
   • Spot-check one row and one percentage. Pick a region: its revenue, then its growth or margin, worked out yourself. fmi_4_1
   • Check the scope. Right quarter, actuals not forecast, and anything you said to exclude (intercompany, returns) actually excluded. fmi_3_1
   • Look for figures that can't exist. Something new this period has no growth rate. It should say n/a, not a number. fmi_1_1
   • Check renamed or merged categories. A region, product or account that changed name should still be matched to its target. fmi_1_5

   With code: the checker the labs used

   Compute the figures yourself, then let the SDK find each one on the slides and list every number that has no source. It flagged every wrong deck in the messy-export lab.

   pip install "git+https://github.com/npmav18/modelcensus#subdirectory=sdk/python"
   
   from modelcensus_toolkit import checks
   from modelcensus_toolkit.checks import Figure
   
   expected = [Figure("Q3 total", 12_010_900, "money"),
               Figure("Midwest growth", 11.46, "pct", "Midwest")]
   for r in checks.reconcile(deck["slides"], expected):
       print("✓" if r.ok else "✗", r.figure.name, r.found)
   print(checks.unsourced(deck["slides"], expected))

   SDK on GitHub →