AI-native building tools at Microsoft AI

1/ A few weeks ago in Microsoft AI, we launched AI-Native building tools: ideas become grounded-prototypes, everyone can collb on, with a path to production, in a fraction of the time.

2/ The Kit
The model was never the bottleneck. Everyone has access to the same models. The gap is context: product strategy, design system, content standards, critique heuristics; without a central knowledge base, you're looking at DIY chaos, design drift and inconsistent quality.

3/ So we built a kit that brings in our standards, context and skills into one repository. We bundled the best mds that generated the best outcomes. We added MCPs & agents for more complex tasks (components code; full content scrub. Quality went high and more teams joined!

4/ We evolved our kit and went deep on token usage, latency, ui consistency and output quality. (And frankly, this is money on the floor – benchmarking models is nice; benchmarking models on your own harness, that's hard). The learning compounded every round.

5/ Now, all of this was in VSC.. An IDE that was never built for designers, or PMs, So we built Studio – "Lovable for (y)our product; grounded in product truth". A desktop app leveraging CLI and existing models/quotas. No per-seat subscription; No tokens markup.

6/ So much was baked into it - templates, to never start from scratch; Plan mode, as one-shot hardly works; Point-to-prompt so you don't need so many words to describe an element on the page; and code vs visual view to verify models use real components for eng to lift and shift.

7/ This made quality and speed go up. Designers and PMs can execute on ideas fast and grounded. Then we took it multiplayer. First, we built Vibehub for people to host their prototypes and be able to share them with their peers. It grew beyond imagination across all of Microsoft.

8/ Then we connected it all - you can upload to Vibehub directly from Studio (multiple versions as you iterate) and share the live demo from Vibehub for teammates to see.

9/ Then we went multiplayer - we added git and remix functionality and commenting (that turns into prompts). One can now see a project in vibehub > hit remix > and colab. This felt transformational.

10/ What we've learned - you can achieve a lot with building 1P software nowadays and you don't need a ton of people for it. We built this with a part time PgM, a part time UXD on MCP, a part time UXD on Kit and a part time UXE (that has decades of experience so that helped :D)

11/ We've also learned that different personas use these tools very differently. The non-product folks want an idea fast. They can use Sonnet and for them the model should not value consistency as much as other personas or use cases.

12/ Then you have the practitioners - the UXDs and PMs that use these tools as their workhorses - they use Terra / Opus models and they appreciate and demand for the harness to overvalue consistency and adherence to platform design rules and product truth.

13/ Last - you have the execs. They overload the task with context, want to break the mold and show something innovative and inspire their teams, and don't have time to babysit the agents. They need the SOTA models, looser guidance and expect high-quality in one-shot.

Fin/ Thanks to @lucascolusso for starting all of this with the kit, @Jeeyoungjung for the all the design system foundations, Ankita Dasgupta for beating the drum, Max Parola for MCP and Eugene Gavriloff for the insane talent and magic in bringing it all together.