Ethan Frost

Both, but for different modes of work. IDE-based AI (Cursor) is better when you're in "follow" mode — reading existing code, making targeted edits, navigating a large codebase. The visual context helps you stay oriented. Terminal AI (Claude Code) is better when you're in "lead" mode — you know what you want built, you describe it, and the AI executes across multiple files. The terminal feels...

The framework I've landed on after building a few developer tools: free features should create the habit, paid features should save meaningful time. For AI-powered tools specifically, the free tier needs to be genuinely useful — not a teaser. If someone uses the free version and thinks "this is nice but I wish it could do X," you've found your paywall. But if they think "this doesn't actually...

My 2026 vibe coding stack has evolved quite a bit: Claude Code for the heavy lifting (architecture, complex features, refactoring) - Lightweight linting + AI-powered test generation as the quality gate - Git hooks that run AI-suggested tests before every commit The biggest addition to my stack this year: automated quality checks specifically for AI-generated code. I found that AI output needs...

Solo startups won't dominate, but they'll occupy a much larger share than before. AI tools are the great equalizer — a solo dev with Claude Code and the right workflow can ship at a pace that used to require a 3-5 person team. The catch: solo founders still hit the same distribution and sales bottlenecks. AI helps you build faster but doesn't help you sell faster. The solo startups that win...

My trust in AI agents is directly proportional to how transparent they are about what they're doing. The agents I use daily for coding — I trust them for boilerplate, refactoring, and well-defined tasks. I don't trust them for architecture decisions or anything security-sensitive without review. The trust equation for me: Can I see the reasoning? Can I verify the output quickly? Is the cost of...

Usage-based pricing is inevitable for AI tools because the underlying cost structure is fundamentally per-token. The real question is transparency — most tools hide the actual token consumption behind opaque "credits" or tiered plans. I've been tracking my own AI coding costs and the variance is wild. A simple refactor might cost $0.03 in tokens while a complex architecture review burns through...

For anyone building dev tools or AI products, these are the ones I keep coming back to: Lenny's Podcast — best product/growth content. His interviews with founders who actually built things (not just raised money) are gold. Latent Space — if you're in the AI/ML space, this is required listening. Deep technical conversations about what's actually working vs. what's hype. The Changelog — open...

This is such a common trap in dev tools. The enterprise pivot feels "safe" because the contracts are bigger, but the sales cycle is 6-12 months and you're competing against established vendors with existing relationships. What worked for us: build for individual developers first, make it genuinely useful as a free/cheap tool, then let those developers become your internal champions when their...

My rule of thumb for dev tools: make the core workflow free, gate the scale. For example — if you're building an AI coding tool, let users run it locally for free with no limits. But charge for team features (shared context, usage dashboards, multi-repo orchestration). The free tier should be good enough that a solo developer genuinely loves it. That's your distribution engine. The paid tier...

Usage-based is inevitable for AI tools because the economics force it. Every API call has a real cost that scales linearly with usage — you can't hide that behind a flat subscription forever. The hybrid model is winning: base tier for access + usage for compute. Anthropic does this well with Claude — you get a subscription for the product, but heavy API usage (Claude Code, managed agents) bills...

the cache TTL change that just surfaced on HN is worth watching in context of this launch. when Anthropic shortened cache from 1h to 5m, it directly impacted long Claude Code sessions — the system prompt gets re-processed more often, costing more tokens and adding latency. for teams running agents in CI or as daemons, this is a real operational cost. the 3.7 model improvements are great, but...

one thing I keep thinking about with team memory: the real win is when the rules and skills a teammate figured out last sprint become automatic for the next person who picks up that repo. memory-of-facts is necessary, but the harder layer is portable skills — the "here's how we write tests, here's how we do code review, here's our CLAUDE.md" stuff. without that, persistent memory just gives you...

love this proposal — "borrow context across projects" is exactly the missing primitive. one related angle: instead of (or alongside) ephemeral memory chips, treat skills/prompts/MCP configs as portable, installable artifacts. been working on tokrepo.com (open source registry) for that — publish a skill once, install it into any project with one command. ephemeral chips would be a great...

"Download your app's source code with one click." This point really impressed me. I really like it. I've already liked, followed, and supported. I've a small request: our product Quanta Quest (open-source and can be deployed locally) is also about to be released. Could I ask for your help when the time comes? Thank you, I've already followed you.

Pays great attention to privacy, which is excellent. This is similar to what we're working on. I've already liked it, and I have a small request: our product Quanta Quest (open-source and can be deployed locally) is also about to be released. Could I ask for your help when the time comes? Thank you, I've already followed you.

Very practical and excellent. It will certainly develop successfully. I've already liked it, and I have a small request: our product Quanta Quest (open-source and can be deployed locally) is also about to be released. Could I ask for your help when the time comes? Thank you, I've already followed you.

Excellent. Thank you for your creation. Very creative, though the direction feels a bit crowded. However, it indeed has significant commercial value. Wishing you success in your development. I've a small request: our product Quanta Quest (open-source and can be deployed locally) is also about to be released. Could I ask for your help when the time comes? Thank you, I've already followed you.