Unproductive Dreams

I gave myself permission to be unproductive yesterday.

I have been juggling local Claude Code agents for a handful of different projects, running simultaneously, and switching back and forth as they need attention. I've been working on some strategies to improve my workflow here, surfacing the agents that need approval or are waiting for instructions - but that's productivity. We're here to talk about being unproductive.

Specifically, imagining my agents at work!

To do this, I'm taking a description of the project, the last 10 git commits, and a little more context from my notes of what I'm currently working on. This gets piped into a text model (currently, Gemini 3 Flash Preview) to generate a structured object:

export const sceneDescriptionSchema = z.object({
 pose: z
  .string()
  .describe(
   "The employee's physical pose and expression, e.g. 'hunched over laptop, brow furrowed' or 'leaning back with arms behind head, grinning'",
  ),
 attitude: z
  .string()
  .describe(
   "One or two words for overall mood/energy, e.g. 'exhausted', 'triumphant', 'frantic'",
  ),
 items: z
  .array(z.string())
  .describe(
   "Props the employee is actively interacting with, e.g. ['towering stack of sticky notes', 'half-empty coffee mug', 'rubber duck']",
  ),
 roomDetails: z
  .array(z.string())
  .describe(
   "Notable details about the room that reflect project status, e.g. ['whiteboard covered in crossed-off tasks', 'sunlight streaming through window', 'overflowing bookshelf']",
  ),
})

If the git log shows no activity for a while, the agent's pose reflects this, and he will appear less and less engaged in his work.

This gets turned into a prompt for Seedream 4.5:

function buildPrompt(scene: z.infer<typeof sceneDescriptionSchema>): string {
 const items = scene.items.join(', ')
 const room = scene.roomDetails.join('. ')
 return `A young man ${scene.pose}, looking ${scene.attitude}, surrounded by ${items}. ${room}.`
}

Then the prompt and a couple of reference images (themed for the specific project) are piped into Seedream to generate an image.

Then the images for the projects are just collected into a web page and deployed to Cloudflare Pages, AI image artifacts and all.

Doing this for a single project isn't a big deal. Seedream costs about four cents an image. But multiply that times several projects, and if I'm updating the photos frequently, it adds up. So, I don't know that I'll keep this running indefinitely, but it's a fun experiment while it lasts!