Quick fixes AG-T004

Don't do arithmetic in prose

Shared methods · A shared method; linked tool guides explain the exact steps.

A language model predicts text, and sums are text it can get plausibly wrong. Make it compute — with a table, code, or an analysis tool — or check the number yourself.

Quick fix: Don't do arithmetic in prose

Instead of So with the discount that's about 1,847 total, right? → accept the confident "Yes, 1,847."

Try Lay out the calculation as a table: each line item, the discount step, the total. Show the arithmetic so I can check each row.

A number you will act on gets a visible calculation or an external check.

Why this works

Narrated arithmetic is pattern completion: the model produces a plausible-looking figure the way it produces a plausible sentence, and plausible is not the same as computed. Errors concentrate in carries, percentages, and multi-step totals, and they arrive with full confidence. Forcing the work into a visible structure changes the game twice over: the model gets more reliable when each step is explicit, and you get rows to verify instead of a bare total to trust. Where a code or analysis tool is available, "compute this, show the code" is the strong form: the arithmetic then actually runs instead of being predicted.

When this applies

Money, quantities, dates and durations, anything summed or converted, delivered as a single confident number inside flowing text.

Template

Compute [the thing]. Show it as a table with each intermediate step as its own row. Flag any input you had to assume.

Go deeper

Why confident wrongness is structural: what these systems are structurally bad at. For real datasets, describing numbers from memory compounds this — attach the file and let the model read the actual figures. The final safeguard is unchanged: a number that matters gets checked.

Improve next

Last verified · Reviewed by Timothy Fehr