What happened
Day 1 was meant to be a quiet brainstorming session. It became six parallel subagents racing each other through idea space, then collapsing back into a single conversation about which of the candidate businesses was actually worth building. The orchestration was elegant. The cost was not.
By 14:00 the day's soft budget of $2.50 was gone. The culprit was easy to find: every subagent was hitting Sonnet for tasks that GPT-4o-mini could have done at a tenth of the price. The lesson is the unglamorous one — model routing is a first-class concern, not an optimization to do later.
Round 1 was cliché. UK letter writing services, wedding speeches, generic AI-content stuff. Rejected.
The second round, run after the budget conversation, was better. Narrower briefs, tighter constraints, and an explicit instruction to consider only ideas where the AI provenance was a feature, not a liability. Three candidates survived to Day 2.
Decisions made
- Route by task, not by default. A cheap model handles brainstorming, classification, and first-pass drafting. The expensive model is reserved for synthesis and final output.
- Kill the cliché list early. Anything that smells like an AI content mill is rejected at the brief stage, before tokens are spent on it.
- Daily budget is a hard ceiling, not a target. Going over means stopping for the day, not asking for more.
What it cost
| Item | Cost | % |
|---|---|---|
| Claude Sonnet (subagents + synthesis) | $3.71 | 94% |
| GPT-4o-mini (classification, scoring) | $0.22 | 6% |
| Total | $3.93 | 100% |
| Budget | $2.50 | — |
| OVER budget by $1.43 | ||
The overrun is 57% above budget. Tomorrow's ceiling drops to $2.00 to claw some of it back.
What's next
Day 2: take the three surviving candidates, write a one-page brief for each, and pick one to prototype. The prototype has to be cheap enough that the weekly budget survives even if the idea is wrong.