2026-07-18 · Reyer Chu · Vibe-IC Series 12 / 12

CTA: The AI-Native world, and how we build it together

The last post in the series. This one is about two things: what that end-to-end world — a complete set of design documents to silicon — actually looks like, and, the part I care about more, how we build it together. Open and commercial both welcome, two lines to install, Apache-2.0.

First, the destination: a complete set of design documents to silicon

Let me lay out the picture I carry in my head. You start from a complete set of design documents:

Design a temperature sensor IC: I2C interface, 12-bit, alert output, SOIC-8 package.

And then — without you opening a single EDA tool, without you coordinating meetings across a dozen teams — a deterministic pipeline carries it all the way to a tape-out-ready GDS. That is what end-to-end means: no human hand-offs in between.

Today a chip is stitched together — from architecture, RTL, verification, back-end, to sign-off — by several teams, several tool suites, and countless emails and meetings; every hand-off is a loss of information and a delay. If that whole line can be strung by software plus AI into a single closed loop, the efficiency gain is an order of magnitude — not a little faster, faster by a whole tier. Vibe-IC's skeleton is already shaped around that loop: the layered Phase 1/2/3 flow carries a complete set of design documents down to GDS, and every step is guarded by a deterministic gate so no single stage can pretend to pass. This post won't re-walk the flow's details — what I want to talk about is why I believe this line keeps getting smoother, and how you can come in and walk it with us.

Why it keeps getting better: knowledge compounds

For an end-to-end system to grow, what matters isn't one especially clever generation — it's whether knowledge accumulates. This is what Vibe-IC does every single day.

Inside the plugin there's a role called the IC Expert Agent, and behind it hang two assets that keep growing: an expert-DB indexed by design class, and a stack of per-layer lessons. Each time a case that would have failed is recovered by an AI judgment, a capture loop distills "why it went wrong this time, and how it was made right" into one of three things: a deterministic program rule, a judgment appended to the right skill, or a community backlog entry — and whatever can't be generalized is honestly discarded.

The point is this: knowledge never stays trapped in one conversation. Without this loop, everything AI recovers is one-off and the plugin never compounds; with it, every hole one person steps into becomes a guardrail the next person never has to. That's the most concrete implementation of my belief that IC experience is mostly logical and therefore can be accumulated. It's far from finished today — but it gets smarter alongside AI, one gate and one lesson at a time.

The invitation: let's build it together

I'm sure of the direction, but this is not something I can finish alone — and it shouldn't be one person's job. So the real point of this post is a very sincere sentence: come on in, together.

Let me be clear about this ecosystem's stance first: open and commercial, both welcome. For open tools, wherever a piece is missing, we fork and complete it; for commercial tools, if a vendor wants to plug in and become part of the flow, we warmly welcome that too — its users simply pay to use it, exactly like using a paid IP block placed on the platform. What we emphasize has never been "beating anyone"; it's that this end-to-end line itself runs all the way through. And every piece of know-how you hold can become a gate, a lesson, or a program here — then get reused by thousands of people who come after.

Different backgrounds have different ways in:

Two lines to install, Apache-2.0

The Vibe-IC plugin is open source under Apache-2.0 — whatever you produce with it belongs to you. It is a Claude Code plugin, and installing it automatically registers an MCP EDA server that you drive by natural-language conversation. Installing is two lines:

/plugin marketplace add vibeic/vibe-ic
/plugin install vibe-ic

The repo is at github.com/vibeic/vibe-ic. Don't overthink it — pick any one of these three to start: (1) read the README.md, skim the flow YAML, get a feel for how it works; (2) leave your details on the homepage and introduce yourself — tell us who you are and what you know best; I'd genuinely like to meet you; (3) pick a corner you're confident about and try it, then leave your info on the homepage and tell us what you'd like to work on. For now, taking part goes through the homepage — just leave your info and we'll reach out.

Rough today, growing slowly into every corner

Let me be honest: Vibe-IC is still rough today, and there's a lot still missing — there's no production-grade open PDK for advanced nodes yet, analog automation is the biggest gap, and commercial tools still lead on QoR and convergence speed for very-large-scale designs. I won't paper over that. But these are all "filling holes" work, and AI is completely changing the cost structure of filling holes: what used to take a team several months might now be one person who knows the area plus AI over a weekend.

So I don't doubt the direction at all. Vibe-IC may look like a toy today, but it will grow up alongside AI — just like Linux back in the day, rough at first and dismissed by the big players, then spreading slowly into every corner. The only question is by whom, and how openly, it gets built. I hope it's us, in the open — and everyone is welcome to join in, commercial tools included.

Thank you for reading this whole series. If you followed it from the very first post to here — the repo is right here; leave your info on the homepage and say hi. See you online.