Key Takeaways

  • Reddit Proto reached 100% adoption across Reddit's product and UX teams, with a company dashboard tracking over 700 active prototypes.
  • Visual designers and animators outperformed technical product managers at building AI prototypes because they treat models like film directors instead of writing rigid specifications.
  • The team eliminated setup friction by abstracting Git, connecting live authenticated production data, and deploying custom Model Context Protocol (MCP) servers to map design tokens into Cursor.
  • When LLMs resist instructions and loop into errors, Metcalf avoids micromanaging: he steps away, strips away granular context, and prompts with broader creative intent.

Why Animators Beat Technical PMs at AI Prototyping

Most software companies assume engineers and technical product managers will lead the transition to AI-assisted coding. When Blair Metcalf rolled out Reddit Proto, an internal environment designed to get teams building functional prototypes with large language models, the data proved that assumption wrong.

“We have 100% adoption in UX and product,” Metcalf explained. “But there are some surprising folks that show up. A visual designer turned out to be one of our best prototypers. An animator turned out to be one of our best prototypers.”

Technical PMs often fail with generative models because they treat prompts like bug tickets or technical specifications. They attempt to specify every edge case upfront, get frustrated when the model drifts, and end up wrestling with syntax. Animators and visual designers approach the tool with a completely different mindset. They think in motion, tone, and end states.

“And what it seems like is these people are like great directors,” Metcalf noted. “They have huge context. They can describe all of the things. They are vulnerable and they try things without worry of failing.”

Because they do not protect an identity as a software engineer, designers iterate freely. When a screen breaks, they reset and prompt again. That lack of ego speeds up cycles and yields better software. Metcalf paired that creative freedom with real production guardrails: Reddit Proto abstracted away Git, hooked directly into live authenticated production data, and ran custom MCP servers that fed real design tokens into Cursor. When prototypes look and behave like the actual product, reviews stop stalling.

"It's instant alignment," Metcalf said. “And so if you use systems with high craft, all the velocity comes because you have alignment with leadership using real data, reviewing real design system components that nobody's hanging up on a gotcha and they don't rat hole on a detail, and they can quickly approve it.”

The Under-Prompting Fix When Models Get Stuck

Every builder hits the wall where an LLM turns stubborn, repeats broken patterns, or refuses to correct an error. The common reaction is to double down: add three paragraphs of rules, write angry warnings in all caps, or micro-manage variable names.

Metcalf does the exact opposite. When a model gets obstinate, he stops typing.

“So, I don't yell at them... what I do is I take a day and then I get more vague,” Metcalf said. “And we all have bots with projects and context, and I pop out of a project and I describe the thing large. Like a movie director.”

Micromanagement locks models into narrow, brittle solution paths. When you feed an LLM pages of constraints after a failure, you narrow its search space around the very bug it just created. Stepping outside the chat, clearing the context window, and describing the goal in broad brushstrokes forces the model to re-evaluate the architecture from first principles. By under-prompting, you let the model find a cleaner implementation path that your granular corrections were blocking.

What to Do With This

Audit your internal tooling setup before your team writes another prompt. If your designers still hand off static Figma files while your engineers complain about prompt quality, build an MCP server that injects your production design tokens directly into Cursor. Tomorrow morning, take your most visual team member, set them up with that environment and live test data, and tell them to direct the model instead of handing off redlines.