Alex MacCaw on AI-Powered Note-Taking

Alex MacCaw built Clearbit (sold to HubSpot), and now runs Reflect Notes—an AI-powered note-taking app with networked thinking at its core.

Reflect: Notes with AI Built In

Reflect isn’t another Notion clone. It combines backlinked notes (similar to Roam/Obsidian) with GPT-4 and Whisper integration. The AI isn’t bolted on—it’s woven into the core experience:

• Voice transcription with human-level accuracy using Whisper
• Chat with your notes to find and organize information
• Generate outlines from scattered thoughts
• Custom prompts you can save and reuse
• Meeting notes that pull action items automatically

Everything stays end-to-end encrypted. Even Reflect’s team can’t read your notes.

Vibe Coding for Senior Engineers

MacCaw’s blog post “How to Vibe Code as a Senior Engineer” explains his approach to AI-assisted development. The key insight: vibe coding works better for experienced engineers, not worse.

“If you know what you’re doing, have a deep understanding of the frameworks and libraries, and a clear idea of the way you like to do things, Vibe coding is like adding Nitroglycerin to your productivity.”

His setup:

1. Great scaffold — Start with a monorepo that has rich examples. AI learns patterns from your codebase.

2. Strong rules — Use .cursor/rules to codify conventions. Force the AI to plan before coding, run typechecks after changes.

3. Perfect context — Manually open all relevant files (including .d.ts types) and add them to context. AI doesn’t gather its own context well.

4. Top models only — Claude Opus 4, Sonnet 4, Gemini 2.5 Pro. Always use thinking mode.

5. Audio prompts — Use SuperWhisper to ramble out messy prompts. Longer, natural language prompts often work better than terse commands.

What AI Still Sucks At

MacCaw is honest about current limitations:

• Context management — AI won’t automatically find what it needs
• TypeScript types — Defaults to any too often
• Planning — Jumps into code without confirming approach
• Taste and architecture — Lacks sense of design patterns

That’s where senior engineers still matter. You define architecture, enforce conventions, and keep the AI on a short leash.

WorkGPT: An Agent Framework

Before the current wave of agent tools, MacCaw built WorkGPT—a GPT agent framework for invoking APIs. It’s a glimpse into his thinking: give LLMs access to real-world actions through well-defined interfaces.

On Artificial Knowledge Creation

His essay “Artificial Knowledge Creation” explores whether LLMs can actually generate new knowledge:

“If you subscribe to the idea that creativity is a form of information extrapolation, then I suspect you’re in the ‘LLMs are already creative’ camp. It doesn’t matter that the dataset is static. You can still combine previously uncombined areas of the dataset to generate new knowledge.”

He draws parallels between David Deutsch’s conjecture/criticism theory and how LLM+evaluator systems work (like DeepMind’s FunSearch). The LLM provides conjecture. The evaluator provides criticism. Loop until you find something new.

The Scaffold

MacCaw open-sourced his ai-monorepo-scaffold—the exact setup he uses for vibe coding. It includes:

• Database migrations
• tRPC routes
• Zod schemas
• React Hook Form setups
• Pre-configured Cursor rules

His philosophy: AI performs best when it has clear examples to learn from. Give it a strong scaffold with good conventions, and the output quality jumps.

Links

• Blog
• Twitter
• GitHub
• Reflect Notes
• WorkGPT
• AI Monorepo Scaffold

MacCaw on the current moment: “It’s the best time ever to be coding. The tooling is magical. The models are competent. The velocity is intoxicating. And most importantly—we still matter.”