Alice Alexandra

You need evals to ship AI features AI features are unpredictable and traditional tests fall short. Evals, automated checks for AI behavior, help you prevent regressions and measure success.

If you're shipping AI without evals, you're running prod on vibes.

At Builder, we wire evals into CI, A/B tests, and prod to catch regressions and prove AI feature upgrades actually help our users.

Here's how to set up your own.

www.builder.io/blog/ai-evals

A two-panel meme using stills of the main character from the TV show Squid Game. The top panel shows the character smiling gleefully with the caption, "watching ai generate code." The bottom panel shows the same character with a serious, stressed expression, captioned, "reviewing ai generated code."

Some tips for code reviews without the tears: www.builder.io/blog/code-r...

Test-driven development with AI Learn how AI transforms test-driven development (TDD) from a time-consuming chore into your secret weapon for building robust and bug-free applications.

TDD feels bad. If I know how to write the function, why start with the spec?

But AI agents switch it up. They need guardrails, and tests let them iterate on even the most complex problems.

I'm convinced AI TDD is the way. Here's what changed my mind:

www.builder.io/blog/test-d...

So, does macOS 26 solve the bug where the Passwords app always opens behind your currently open app? I think that's all I care about for advancements.

looking at the sink: "ugh, time to do plate laundry again"

Code review in the AI age AI hasn't replaced developers; it's promoted them to architects. Learn how to review AI-generated code and use tools that create review-friendly diffs.

Reviewing 1000-line AI-generated diffs sucks. Even AI can't really do it right.

Here's how to stop being an AI babysitter and start being a code architect.

www.builder.io/blog/code-r...

Mock up a website in five prompts Learn how to create a website mockup in minutes using AI-assisted tools. This guide walks through building a bakery site from concept to working prototype.

Copy/paste from any website into your own design system with Fusion: www.builder.io/blog/mock-u...

you gotta scrape OpenAI back sometimes

Builder.io: Visual Development Platform Builder is the only Visual Development Platform that offers an AI-powered design-to-code tool, a visual editor and an enterprise CMS.

You can try it out at builder.io/fusion.

Keeping Figma in sync with Storybook (and your entire design system) just got way easier.

Weird anti-AI scraping technique: digital tar pits.

Websites hide links that lead bots to an endless maze of auto-generated gibberish that all link back to itself, poisoning training data.

Wouldn't recommend, but still an interesting tactic against overly eager scrapers.

Figma's *official* Dev Mode MCP server is now in open beta, and it's a pretty awesome tool!

I've updated this article with how best to use it, how its AI works under the hood, and where some alternate workflows can augment your design to code process.

www.builder.io/blog/figma-...

sometimes when i listen to Hozier, i'm just like, "Andrew, stop it, we're in _public_"

A diagram illustrating the MCP architecture, with the title "Pick the right one" pointing to the client. The 'MCP client' is shown connecting to various AI models (like ChatGPT), while it communicates with an 'MCP server' that connects to different backend services (like Figma, Notion, and GitHub).

Hot take: Design handoffs won't work until designers can ship code.

Current MCP clients force you into dev tools or basic chat. PMs spec in docs, designers mock in Figma, and devs rebuild everything.

What if one MCP client let the whole team code? 👇

builder.io/blog/mcp-cl...

If you're interested, you can try it on your own repo, right from the builder.io homepage.

Debugging dark mode issues in an old, messy SvelteKit codebase is usually nightmare fuel.

This morning, I connected my repo and put Builder Fusion on the case. It agentically iterated with the code and full DOM context, and it figured out the problem with a single prompt.

Dictionary entry for "bafflegab," a noun that means "incomprehensible or pretentious language, especially bureaucratic jargon."

new favorite word just dropped

A visual comparison of two development workflows. The top timeline moves from “Idea” through “Wireframe,” “Prototype,” and “Code” to “PR,” ending with a red X. The bottom skips intermediate steps, jumping from “Idea” directly to “PR,” with a dotted line labeled “Time saved” and a green checkmark.

What if you could mock up prototypes that stakeholders can actually click through?

You don't have to just pretend it works.

Use a real GitHub PR workflow and mock a site all in less time than it takes to wireframe. Here’s how.

www.builder.io/blog/mock-u...

Image illustrating Figma Model Context Protocol (MCP) translating a product card design (left, showing visual layers and dimensions) into a React code component structure (right, with placeholders like ProductImage, ProductName).

Figma's new MCP server gives AI access to structured design data instead of screenshots. But is the resulting code pixel-perfect?

I've put it to the test is my latest post.

www.builder.io/blog/figma-...

The word “login” comes from throwing a log attached to a rope with knots overboard a ship to see how many knots go by over time (see also, knots as speed). You’d then put that info in the “log book.” You’d “log in” on a regular basis. This wasn’t from 1959, it was likely from 1689! Etymology baby!

Visual comparison highlighting a preferred method for AI development. The "Vibe coding" approach (disapproved with a red X) shows a direct, simplistic request to "Create an ecommerce app". In contrast, "AI for grown-ups" (approved with a green checkmark) shows a more structured approach involving design (Figma), code, and data considerations feeding into the AI prompt.

My problem with vibe coding isn't the AI; it's the vibe.

Many AI tools disrespect your team's hard work and talent. What we need is AI for grown-ups.

Here's my thoughts on what that means.

builder.io/blog/ai-for...

an AI agent that reminds me why I walked into a room

ugh stop posting such relevant content

I want less girlboss & more girlfail having a very bad day but being very brave representation

You can teach any agent to fish, but wouldn't you rather it know who to call to get fish on demand?

This is what Google's new A2A protocol promises: your agent gets a list of contacts for when the questions get too tough.

www.builder.io/blog/a2a-pro...

AI agents are like Twitter devs: great solo, terrible at collaboration.

MCP gave agents standardized tool access to information across the internet, but the story isn't done.

A generalist agent with tons of tools still isn't as useful as an orchestrated network of specialists.

Yep. I’d rather a hastily thrown together diagram than some six-fingered action figures.

This is dedication. 😁

Teach the model once, and then let it work.

In other words: fine-tuning.

- 50 great examples > 5k-token prompt.
- Tone of voice and tool calls get baked right into the model weights.
- Adapters make this possible for all. No second mortgage for GPUs.

www.builder.io/blog/fine-t...

Diagram showing the composition of a full prompt sent to an LLM, consisting of a large "chonky system prompt" ($0.25 per request) and a small "user query" ($0.001 per request), highlighting the cost difference.

If you're in the midst of yet another 5,000-word manifesto-as-AI-prompt, it might be time to take a beat.

There's a nicer route...