@garrettDoesCode

Creative Developer at ESPN. Design engineering, interactive storytelling, and AI-augmented dev workflows.

Florida
Joined December 2023
Graphic designer turned production engineer at ESPN. I share what I learn building interactive stories, frontend systems, and practical AI workflows—especially the decisions behind the finished work.
2
113
Garrett Siegel retweeted
74
661
83
8,598
254,043
Who here remembers how to write code?
0%Yes
0%No
0 votes • Final results
1
11
ChromaKit shows 941 weekly npm downloads, but that doesn’t tell me what to build next. One GitHub issue gave me a concrete need I could address. Small open-source maintainers: how do you get feedback from people who don’t open issues?
🤖 Made with AI
8
Designers: how would you feel if a teammate used AI to animate part of your finished artwork? Even subtle movement adds a creative decision you didn’t make. Would you want to approve the idea first, see a test, or handle the animation yourself?
🤖 Made with AI
8
I had GPT-6 Astra read its own model guide, then gave it this prompt: “Interview me—grill me—in plain English about how I work, then configure the best setup for my preferences.” It has yielded much better results in my agentic work. developers.openai.com/api/do…
1
57
Most components are reviewed with perfect content: short headline, clean image, exactly three lines of copy. Before handoff, test four cases in design and browser: short / long / empty / error Content isn’t poured into a finished layout. It’s an input that creates it.
12
A webpage is not a poster. It has to survive wrapping, loading, errors, keyboard focus, zoom, touch, reduced motion, and widths you never mocked up. Working as a developer taught me to design rules, not perfect frames. What web constraint changed how you design?
11
Garrett Siegel retweeted
Unpopular opinion: AI is only as smart as the person using it.
210
132
24
1,088
31,136
My AI setup never stays still for long. Some things become part of how I work. Some feel useful for a week. Some just create another system to maintain. What has actually earned a permanent place in your workflow?
5
Hot take: most “AI agents” are prompts wearing a tool belt. If the job is one clear instruction, adding planning loops, tools, and autonomy doesn’t make it smarter. It gives the task more places to fail. What’s the simplest thing you’ve seen overbuilt as an agent?
11
Okay, I get the Pi hype now. Tonight I finally gave it a proper shot and ended up spending hours tweaking the harness. Codex. Claude. OpenCode. OpenRouter. One minimal interface. I’ve already customized it around my workflow and trimmed my system prompt down to ~3k tokens, including all of my current tooling. For those already deep into Pi, what’s one customization you can’t live without?
51
I keep having the same reaction to AI launches: impressive, then immediately familiar. So many people are building toward the same few ideas that a product can feel dated before it ships. Is that real market pressure—or are we all just watching the same feed?
15
The line improving my AI-built code most: “Do not build the whole MCP server at once. Teach one step, explain why, show how to verify it, then pause.” I add TypeScript, the current spec, and a URL to data I am allowed to use. I want ownership, not a mystery repo.
2
25
The MCP learning order that finally helped me: beginner explanation → architecture → current spec → tiny build → explain it back. Starting with the spec made everything feel larger than it was. Building one showed me the core idea: a connector with boundaries.
13
MCP felt much more mysterious than it needed to when I started building one tool for work and another for myself. Once I separated the host, client, server, and allowed capability, it clicked: it is a connector with boundaries. The boundary is the design work.
19
I use Figma for web design and Photoshop for print or heavy image work. Neither one “wins.” They help me answer different questions. The browser still gets the final say on whether the web design actually behaves.
18
I can make a component look finished with perfect content. The real test is what happens while it loads, when there’s nothing to show, when something fails, or when the content refuses to fit. That’s the version the design system has to account for.
🤖 Made with AI
16
A design system can give AI every color, spacing token, and component and still get lifeless output. The rules explain what exists. They don't explain why one choice matters more. The useful part isn't more rules. It's the reasoning behind them.
🤖 Made with AI
19
The technical skills I learned did not make my graphic design experience less relevant. They gave it more places to work. Visual judgment and technical judgment now shape the same decisions. That overlap has become the most useful part of my career.
9
A design tool has to protect its promise. Codebase Rescue passed my tests, but agents still broke its audit-only boundary four ways. I fixed each problem and added tests for the promise, not just the implementation. github.com/garrettsiegel/cod…
12