Process automation
A storyboard job arrives as a handful of loose links: a Drive folder, a script in a document, assets the client sent however they felt like. I built a process where agents handle the mechanical part and draw inside Photoshop via scripting, but nothing moves forward without two approvals from me. And on every round the system studies my corrections so it lands closer on its own next time.
I bring the bot into the very conversation where the job landed.
The bot locates the Drive links, the script and the assets the client sent.
An agent classifies the deliverable: storyboard, one-page PDF or landing page. That decision defines which agents work next.
The agent set for that type reads the script, sorts the assets and builds a draft as a static image.
Nothing enters Photoshop without my sign-off. This is the cheapest gate in the process: correcting direction here costs minutes.
The bot draws via scripting, following the guides for that project type. On landing pages it asks first whether to create a design system or reuse an existing one.
I review the layouts by hand, the way I would review any other delivery.
The bot analyses what I changed in review and absorbs it, so the next round starts closer to the final result.
↑ back to drawing, with the learning applied
It uploads files to Drive, generates the PDF or returns a Figma share link.
I send it, already in the format that particular client expects.
Frames from two real batches. Note that the sign-language window and the caption bar come out already positioned by the guide, identical across every frame in the batch.
The static draft exists so I can kill a wrong direction while it still costs minutes, and the final review stays manual. Automation without a gate produces an entire wrong batch, which is the most expensive failure in this kind of production.
One agent classifies the job as storyboard, one-page PDF or landing page, and only then does the agent set for that type step in, with the guides for that type. Without it, a single generic prompt would have to serve three different products.
After I review, the bot analyses what changed and absorbs it. That is what separates this from a generator: the judgement stays mine, but it stops repeating the same mistakes.
On landing pages the bot asks first whether it should build a new design system or reuse an existing one, because that is a project decision and not an execution detail.
Marketing and lead capture site, live.
The price list as the argument, not as a buried table.
The clinic's best asset was buried in a bullet point.
Two specialties, one page, no decision made.