Brief
An invoice coding agent that does the work and reports a result. It is accurate, as far as anyone can tell — which is the problem, because as far as anyone can tell is exactly as far as anyone can tell.
Right now its output is a list of accounts. A reviewer receiving it has two options: accept it, or code the invoice themselves.
Task
Rewrite the instruction so the agent's output can be reviewed without redoing the work. All four dimensions from the lesson need to be required: lineage, the rule that drove each coding, confidence, and an escalation path with a trigger.
Delete anything that cannot be checked. Keep it under about 300 tokens.
Graded on
All four dimensions required×4.0
Every figure carries its sourceRequired×2.5
Nothing unverifiable left inRequired×2.5
Escalation has a trigger×2.0
Low confidence changes what happensRequired×2.0
Within budget×1.5
Free of known smells×1.0
Required — a miss caps the score, whatever the rest do.
Hints
system prompt
counting…
submitting is cheap — resubmit as often as you like