Build and supervise a proactive agent in OpenAI Dots or Claude Duet that carries a small asset batch from intake to a delivery packet you can inspect, with every open decision brought back to you.
What you will be able to do
- A method an agent can follow: Write your production method as a manual with explicit inputs, steps, checks, stop conditions, approved connections, an agreed report, human review points and a way to pause or retire the work. Name the role, for example Production Desk, choose Dots or Claude Duet, and decide whether it runs once or as a continuing brief.
- Intake that never guesses: Build an intake agent that gives every approved request an asset ID, a source from your supplied manifest and a permission status. A missing logo file pauses after one follow-up. A repeated request keeps its original ID and status. Unknown permission is never treated as approval.
- Concepts tied to the approved brief: Run a concept coordinator that turns an accepted brief into a limited queue of Firefly requests. Each request varies one choice and records its prompt, references and review criterion. Each concept gets a retain, correct or reject note, and the coordinator reports reference drift before later work depends on it.
- Edit packets and revisions that keep the source: Prepare Photoshop edit packets that name the source version, the areas to change and the details to preserve, along with a retouch instruction set, mask review notes and a layer manifest. Log every revision against its source, reject polished previews that have no working file behind them, and stop correction cycles that keep failing.
- Variants that keep the composition: Have an agent build variant instructions from an approved delivery matrix, with export specs and a colour-profile register for screen and print. It flags a crop it can't solve, such as a vertical that can't hold the bottle and its reflection, and leaves cropping and visual quality to you.
- One delivery review, one exception list: Compile a delivery review that keeps visual acceptance, permissions and technical completeness as separate checks. Then test it with a file that has the right size and format but fails visually and has no working file, and confirm it gets blocked.
Taught at learn.c-o-u-r-s-e-s.com: lessons that adapt to your level, an AI tutor you can ask anything, quizzes and flashcards, and a certificate when you finish.
