Credit spend gates: no blind re-rolls
Agentic workflows and custom skills
Gating generation spend, and turning every failed generation into a written rule.
A rejected generation is never simply re-rolled. The start frame is checked first, then the references, then the prompt, and if none of those is at fault, the model itself is switched.
Whatever the fix, it becomes a playbook rule, and the retry still goes through the cost preflight and approval.
At Formative Minds I produced short animated films and game assets with generative AI: Higgsfield for images and video, driven through MCP tool calls, and ElevenLabs for voice and sound, through its REST API.
problem
Generation credits are expensive and reset every month, and a failed generation is easy to re-roll on hope. Re-rolling without knowing why it failed tends to repeat the same mistake, and the same kinds of mistakes kept coming back.
solution
Every generation goes through a gate: a free cost preflight, then my approval before any credits are spent.
A rejected result is diagnosed before trying again, checking the start frame first, then the references, then the prompt. The lesson is written into a playbook of rules that the agents read before the next job.
Locked results are logged in a ledger, and a spend skill reconciles the ledgers against the real balance from the API.
On the last production, none of the six rejected sets was a blind re-roll, 13 seconds of film were cut from existing footage for 0 credits, and the full soundtrack used about 5% of its audio quota. The costliest mistake, about 150 credits, came from ignoring one of the playbook's own rules.