Context

Claire Vo talks with John Bai and Peng Zheng, both designers on the Grok Bot team at SpaceXAI, about how they personally use AI agents (bots) to run their creative and production work. This matters to PMs because it's a live demonstration of a pattern worth generalizing well beyond design: turning an ambiguous personal idea into a working pipeline without first writing a spec, and prototyping product ideas cheaply enough that they never have to clear a product-priority gate before someone can see if they're worth pursuing.

The Big Idea

When building something is nearly free, you can start with a vague idea, hand it to an agent, and let the system reveal itself through iteration, rather than needing to know your full requirements up front.

Peng built an entire self-updating personal website, with image processing, a location lookup pipeline, and structured data, starting from a single photo and no spec, Figma file, or plan. The lesson for PMs is that when execution cost drops enough, the traditional gate of "is this worth speccing and prioritizing" becomes optional for a whole category of experiments.

Key Insights

A vague idea can become a real pipeline with no spec

Peng describes starting his personal website project with a single 3D-rendered photo of a Chinatown grocery store and "no Figma file, no spec," not even knowing whether he'd need a CMS or hosting. He just told Grok Bot the idea and asked for the simplest way to build it, then kept adding capability as new ideas occurred to him. The end result, a check-in pipeline that takes a photo or place name, geocodes it via the Google Places API, generates a stylized 3D image, and publishes it, emerged entirely through iteration rather than upfront design.

Removing friction is what actually keeps a habit alive

Peng frames his check-in site as solving a real behavioral problem: he wanted a record of places and people he's spent time with, but every prior check-in app he'd tried eventually got abandoned because logging required manual effort. His fix was to make updating the site as low-friction as sending a photo or a screenshot of a friend's social profile to a bot, since Claire notes that requiring him to "download your photo and copy-paste the profiles" the old way would mean the site would stop being maintained within weeks.

AI unlocks skills a designer never had, not just speed on skills they already have

Claire describes herself as strong on UI/UX craft (corner radii, kerning, layout) but unable to illustrate, and both John and Peng describe using AI (specifically 3D-render generation and motion work) to add capabilities that were previously inaccessible to them without hiring a specialist. Peng frames the shift as a drop in the cost of execution: "the cost of making things become so low that we can imagine something that cost a lot more in the past" and now build it directly.

Production design work can happen while you're physically away from your computer

John built a bot ("Figma Bro") connected to Figma via its MCP server that handles placement, scaling, and styling tasks by voice. In one live example, he directed a batch of icon placements across a Figma file entirely through a voice memo recorded at the gym, then returned to three finished output options already matched against a colleague's Slack request. He frames this explicitly: "I use AI to get me off my computer, not on my computer," positioning voice-directed agent work as freeing him from needing to be at a keyboard for production tasks.

The trash can method removes the gatekeeping tax on trying ideas

John names a pattern he calls "the trash can method of software development": building things on the presumption that discarding them is essentially free. He contrasts this with the historical path a spontaneous idea (his example: a screenshot-style share animation) used to take, getting stopped by a PM asking for ROI or an OKR, or an engineer citing tech debt, and even if it got built, facing resistance to killing it later because "you don't deprecate any of these things" once resources were spent. Because a bot ("DevBot") can turn a described interaction into a working prototype cheaply, John can explore multiple directions and simply not ship the ones he doesn't like, without ever needing organizational buy-in to try.

Bots can be organized into a life-and-work bot ecosystem

Peng organizes his personal bots into life and work buckets, with a "chief of staff" bot he delegates ambiguous tasks to (it then routes and even chains tasks, like buying 3D-printing filament and separately updating his Notion inventory afterward), plus dedicated bots for email triage, calendar scheduling from a screenshot, a personalized 7am news digest, and tracking a government waitlist status monthly. He also runs bots in a group chat structured as a lightweight cross-functional team (a PM bot, designer bot, and engineer bot) that debate a new idea from their respective perspectives once he seeds it.

Mental Models & Frameworks

The trash can method of software development

A mental model for evaluating whether an idea is worth trying: if building and discarding something is close to free, evaluate ideas by curiosity and interaction quality rather than by whether they can survive a prioritization gate. Use it when a team's usual filter (ROI justification, OKR alignment, engineering bandwidth) is blocking cheap, reversible experimentation rather than protecting genuinely scarce resources; the model only holds because prototypes built this way are explicitly disposable and never carry the sunk-cost pressure to ship regardless of quality.

Bot-as-pipeline instead of bot-as-chat

Both Peng and John treat their bots less as a conversational assistant and more as an orchestration layer stitching together plain building blocks: Peng's check-in bot chains the Google Places API, image generation, and structured data output into one flow triggered by a single photo; John's Figma Bro chains Figma's MCP connection with layout and styling rules he's pre-taught it. Use this model when scoping a new automation: identify the discrete, ordinary tools already available (an API, a generation model, a data store) and let the agent be the connective tissue between them, rather than trying to build one bespoke system from scratch.

Trade-offs & Nuance

Cheap iteration still needs upfront constraints to avoid a messy result

John notes that his first Figma Bro attempt needed several rounds of manual direction, while a later attempt worked more precisely because he'd pre-taught the bot specific layout conventions (artboards spaced 50 to 100 pixels apart, consistent orientation) up front. Lowering the cost of iteration doesn't remove the need for the human to encode real constraints; skipping that step just shifts the cost from planning time to cleanup time.

AI expands a designer's craft rather than replacing its purpose

Claire pushes back on the framing that AI threatens creative work, arguing "your highest calling was never to design gradient fills in Figma." Her position is that offloading the lower-leverage production tasks (asset polishing, resizing, repetitive styling) to a bot lets a designer spend more time on the parts of the craft that are actually hard to automate: taste, sound, video, and the specific texture of delight in an experience, rather than eliminating the need for a designer's judgment.

Practical Application

Start an ambiguous idea with the bot, not a spec

When you have a rough concept you're unsure how to build (Peng didn't know if he needed a CMS, image hosting, or a database), describe the end goal to an agent and ask for the simplest path forward rather than speccing the full system first. Let the shape of the solution emerge through a few rounds of "what if it also did X," and only formalize it once it's working.

Pre-teach your agent your working conventions once

Before delegating a repeatable task (John's Figma layout work, for example), spend one session explicitly telling the agent your standing preferences: spacing, naming, formatting, or structural rules you always want followed. This upfront investment is what let his second Figma Bro request come back nearly finished, versus his first attempt needing several rounds of correction.

Chain a personal assistant bot's tasks together explicitly

If you delegate a task to an assistant bot, also tell it the follow-up action that should happen automatically once it's done (Peng's example: buy the filament, then update the Notion inventory record). Teaching the bot the full task chain once means you can later just say "do the thing we did last month" instead of re-specifying every step.

Prototype interaction ideas as disposable, not as scoped commitments

When you have a spontaneous idea for an interaction or feature, describe it to a prototyping bot and let it build a working version without first running it through a prioritization process. Treat the output as something you're free to discard if it doesn't feel right in practice, which removes the pressure that normally keeps small, uncertain ideas from ever getting tried.

Bottom Line

When building and discarding a prototype costs almost nothing, the right response to a vague idea is to hand it to an agent and iterate toward a working system, not to write a spec first; the same shift that lets a designer generate a 3D-rendered personal website from a single photo also removes the organizational gate that used to kill promising ideas before anyone could see them work.

Tools & Products

People to Follow

John Bai

Designer on the Grok Bot team at SpaceXAI who wrote the widely shared piece "Designing Grok Bot with Grok Bot" and publishes bot templates (including "Figma Bro") for other designers to reuse and build on.

Peng Zheng

Designer on the Grok Bot team at SpaceXAI whose personal website doubles as a live demo of his check-in bot pipeline, turning a photo or place name into an automatically updated, stylized entry on his site.