When Taste Becomes the Product Skill, and a Year of VPS Coding

Lenny Rachitsky / Andrew Ambrosino — The New Shape of Product Work When Software Gets Cheap#

Source: https://www.lennysnewsletter.com/p/openai-codex-lead-on-the-new-shape Credibility: High (first-person practitioner interview, Lenny's newsletter, Andrew Ambrosino leads Codex at OpenAI)

What happened: Andrew Ambrosino, who leads product for OpenAI's Codex, joined Lenny's podcast to talk about what changes when software becomes dramatically cheaper to build. The core argument: as the cost of execution drops toward zero, the scarce resource becomes judgment — knowing what to build, what's worth building, and what "good" looks like. That's a product and taste problem, not a technical one. He also discussed what's next for the Codex desktop app and how his team thinks about AI-assisted software development as a product category.

Key patterns:

• When execution cost drops, taste becomes the bottleneck. The limiting factor shifts from "can we build this?" to "should we build this, and will it be good?"
• This isn't just a philosophical point — it has hiring and team structure implications. If taste is the scarce resource, you want people who can evaluate quality, not just ship quickly.
• Codex desktop app is a live product lab: the team building tools for AI-assisted development is also using those tools to build those tools. Tight feedback loops.
• The framing of "software getting cheaper" is more precise than "AI is changing development" — it's specifically about cost curves, and what happens to products and teams when those curves shift.

Why it matters for PMs: The "taste matters more" argument has teeth when you think about what PMs actually do. If engineering capacity expands dramatically (via AI coding tools), the bottleneck in your product org doesn't go away — it moves upstream to product judgment. PMs who can define quality, recognize it quickly, and make confident calls about what's worth building become more valuable, not less. This also has implications for how you staff and evaluate product teams going forward.

Critical questions:

• Is "taste" actually a trainable skill, or is Ambrosino describing something that's hard to develop post-hire? If it's mostly innate, this has real implications for sourcing.
• How do you measure taste in a product org? Great PMs have always claimed this matters — what's different now is the stakes of getting it wrong are higher when you can ship 10x more.
• Does cheaper software mean more experiments (good) or more mediocre product surface area that's costly to maintain (bad)? The podcast seems to assume the former.
• For Codex specifically: how does the team avoid the feedback loop problem of building AI dev tools using AI dev tools? Dogfooding is powerful but can create blind spots.

Action you could take today: Pull three recent product decisions your team made and ask honestly: was the limiting factor "could we build it?" or "did we know what good looked like?" If it's mostly the former, you may be systematically underinvesting in the taste/judgment layer.

Pieter Levels — A Year of Coding on a VPS with Claude Code#

Source: https://levels.io/coding-on-vps-with-claude-code-for-a-year Credibility: High (first-person practitioner account, Pieter Levels is a solo founder with real shipped products and public revenue)

What happened: Pieter Levels published a reflection on almost a year of coding almost entirely through Claude Code running on a VPS, rather than using a local IDE. This is a real workflow, not a thought experiment — Levels has been shipping products throughout this period. The post covers what works, what breaks, and how his development process has changed.

Key patterns:

• VPS + Claude Code as a primary development environment removes the local machine dependency. Your dev environment lives in the cloud and you interact with it through the model.
• Levels describes this as a fundamentally different relationship with code — more like directing than writing. He's specifying intent, reviewing output, and iterating, rather than typing syntax.
• The workflow is particularly well-suited to solo builders: low coordination overhead, fast iteration, no context-switching between tools and teammates.
• A year is a meaningful time horizon. This isn't a weekend experiment — it's evidence that the workflow is stable and productive enough to sustain a real business.

Key technical details:

• VPS (Virtual Private Server) rather than local machine means the model has persistent access to files, terminal, and running services — no handoff friction
• Claude Code operates with enough context to understand project structure over time, not just single files
• This pattern sidesteps local environment setup entirely, which is often the highest-friction part of starting a new project

Why it matters for PMs: Levels is the clearest signal for where solo and small-team building is heading. His workflow is reproducible by anyone with a VPS and Claude access today — it's not waiting on future capabilities. For PMs thinking about AI-assisted development: this is what the "vibe coding" pattern looks like at one-year maturity, from someone who's been doing it continuously with shipped products. The workflow removes a lot of the tooling and environment overhead that makes AI coding tools feel high-friction early on.

Critical questions:

• How does debugging and error recovery work in this setup? When something breaks at 2am, can you actually diagnose and fix it without dropping into a traditional IDE?
• What kinds of products and code complexity does this hold up for? Levels builds relatively self-contained consumer apps — would this workflow survive a larger codebase with multiple contributors?
• Is the VPS pattern better because of Claude Code specifically, or because persistent cloud environments are just better for AI-assisted development regardless of which model you use?
• What's the cost structure? VPS + Claude Code API usage at a year of sustained development is a real expense — is this actually cheaper than traditional tooling?

Action you could take today: If you haven't tried Claude Code (or Cursor in agent mode) on a project with actual complexity, set up a VPS or cloud dev environment this week and run one real task through it. One real session beats a dozen demos for calibrating your intuition about where AI coding tools actually are.