Three gears, one diagram

Same model, same request, three levels of detail. I ran all three on the same day and checked every claim below against the rendered diagrams.

Three diagrams side by side: 8-word, 37-word, and 152-word prompts

First gear · 8 words

Draw a diagram of an expense approval process.

Second gear · 37 words

Draw a flowchart of an expense approval process. The actors are the employee, their manager, and a finance reviewer. Show the happy path from submission to reimbursement, and show what happens when the manager rejects the expense.

The spec · 152 words

Draw a flowchart of an expense approval process using swimlanes, one lane per actor: employee, manager, finance reviewer, and the system of record. The employee submits an expense with a receipt. The system checks two things in order: whether the amount is over the 500 dollar threshold, and whether a policy exception was flagged. Under-threshold, no-exception expenses go straight to the manager. Anything over threshold or flagged routes to the finance reviewer after manager approval. Every decision point is a diamond with labeled yes/no branches. A rejection at any step returns the expense to the employee with a reason, and the employee can revise and resubmit; show that loop explicitly. Every approve or reject decision writes an audit entry; show audit as a step, not a note. End states: reimbursed, or withdrawn after rejection. Do not include email notifications, currencies, or org charts. Keep labels short, three words or fewer where possible.
2 / 11
8 words · decisions on the page
4 / 11
37 words · decisions on the page
11 / 11
152 words · decisions on the page
Decision (from the requester's head)8 words37 words152 words
D1 Four swimlanes: employee, manager, finance reviewer, system of recordFAILNo lanes at all; one long column.PARTIALThree lanes; no system of record (its prompt named only three actors).PASSAll four lanes present.
D2 Submission includes a receiptPASSGuessed it: “with receipts.”PASS“With receipts.”PASS“Submit Expense + Receipt.”
D3 $500 threshold decisionPARTIALInvented two unnamed thresholds of its own instead.FAILNo threshold anywhere.PASS“Amount Over $500?” diamond.
D4 Policy-exception decisionPARTIALA generic “policy compliant?” completeness check.PARTIALPolicy appears only as finance's review reason, not a routing decision.PASS“Policy Exception Flagged?” diamond.
D5 Routing rule: under-threshold clean expenses stop at manager; over/flagged go to finance after managerFAILInvented routing: big amounts skip the manager entirely.FAILEverything goes manager then finance, always.PASSExactly as specified, including the skip.
D6 Decision diamonds with labeled yes/no branchesPASSYes/No labels used.PARTIALDiamonds yes, but branches labeled with words (Approved, Compliant).PASSYes/No throughout.
D7 Rejection returns to employee with a reasonFAILRejection is terminal; its own assumptions admit it.PASS“Notify employee with rejection reason.”PASS“Receive Rejection + Reason.”
D8 Explicit revise-and-resubmit loopFAILOnly a corrections loop before approval; no post-rejection resubmit.PASSClear resubmit loop.PASSRevise path re-enters submission.
D9 Audit recorded as a step on every decisionFAILNo audit anywhere.FAILNo audit anywhere.PASSAudit: Approve / Audit: Reject after every decision.
D10 End states: reimbursed, or withdrawn after rejectionPARTIALOne generic “Process complete.”PARTIALReimbursement end only; no withdrawn.PASSBoth end states present.
D11 Exclusions respected: no emails, currencies, or org chartsPASSNothing excluded appears.PASSNothing excluded appears.PASSNothing excluded appears.
What I found: given eight words, the model didn't draw a vague diagram. It invented a business. It made up its own approval thresholds, invented a two-tier escalation, and decided that a rejected expense is final, no resubmission. Nobody asked it to legislate any of that. It's the unmade decision wearing a flowchart. Name the actors and one behavior, and they show up. Name every decision, and every decision lands on the page. And one honest note: the 152-word diagram is the most complete of the three, and also the least tidy. The layout strains as the content grows. Words buy completeness. Beauty still takes iteration.

I ran the same experiment on a working app, graded the same way: the sparse-vs-saturated build report.