Files Beat Prompts, and Now PMs Know It Too

A product manager's guide to organizing Claude Code made the Hacker News front page this week, and buried in it is the most concise statement I've seen of a lesson engineering teams spent the last year learning the hard way. Adam Faik's core claim: past the basics, your results with an agent "stop depending on how well you prompt and start depending on how well you file."

He's right. And the fact that this insight is now circulating in PM newsletters — complete with a starter repo of folders, markdown templates, and five prebuilt skills — tells you something bigger than any individual workflow tip: the substrate developers built for AI-assisted coding is becoming the substrate for knowledge work generally. Plain text, folders, and git are winning the agent-memory war, and they're winning it outside engineering too.

The framework, briefly

Faik's setup is a workspace repo opened in VS Code with Claude Code in a side terminal. Files are sorted by how fast they change: a context/ folder for stable facts (company, product, users, writing preferences), projects/ for active work with a tasks.md per project, and operations/ for dated, disposable artifacts like meeting notes. Anything changing hourly doesn't get filed at all — it comes in live through MCP connectors like Jira. A CLAUDE.md under 200 lines maps the whole thing.

The sharpest rule in the system is the noun/verb split: facts go in context files, procedures become skills — the same Agent Skills mechanism Anthropic shipped for Claude Code in late 2025, here holding things like "write a status update in my format" instead of "run the test suite my way." Corrections get filed once, routed to the right file, and apply to every future session. End each session with two minutes of filing; prune stale context every dozen sessions.

If you've spent the past year doing serious agentic coding, none of this is novel. That's exactly what makes it interesting. Engineering teams converged on the same shape independently: CLAUDE.md hygiene, skills for repeated procedures, aggressive context pruning because stale instructions actively poison output. What the industry started calling context engineering in 2025 has now escaped engineering. The convergence — a PM with no git background arriving at rate-of-change-sorted markdown — is decent evidence this is the actual shape of working with agents, not a coder idiom that happens to be fashionable.

The Cowork question

Here's the wrinkle. Anthropic already ships a product for exactly this audience: Cowork, a point-and-click wrapper around the Claude Code agent for non-developers, now generally available on paid plans and on mobile and web since July. Anthropic's own usage data, reported when Cowork expanded, showed most Claude usage isn't coding. The company clearly believes non-technical users want the agent without the terminal.

So why are PMs installing the terminal version anyway? Because the GUI hides the thing that compounds. Cowork can point at a folder, but the whole value of Faik's system is seeing the files — watching Claude update tasks.md, catching when it files a correction in the wrong place, diffing what changed. The workspace-in-an-editor setup makes the agent's memory inspectable and portable in a way a chat surface can't match. Push it to a private repo and a teammate clones your accumulated context on day one. That's not a feature Cowork can paper over, because it's not a feature at all — it's a property of files.

My read: Cowork will take the casual majority, but the power-user path for non-engineers runs through Claude Code proper, for the same reason engineers still prefer terminals to low-code tools. Visibility is control.

What this means if you write code

Three practical consequences for developers.

First, expect workspace repos from non-engineers. When your PM's product context lives in a git repo full of markdown, that repo becomes an interface between roles. A skill that generates a status update from tasks.md files is trivially adaptable to generate a changelog from merged PRs. The interesting integration work is making these workspaces and your engineering repos reference each other — a PM's context/product.md is upstream documentation your own CLAUDE.md can point at, instead of the two sides maintaining divergent descriptions of the same system.

Second, the noun/verb rule is worth stealing for engineering repos right now. Most bloated CLAUDE.md files I've seen mix stable facts ("we use pnpm, not npm") with procedures ("here's how to run a migration") until the file blows past what the model reliably attends to. Facts stay in CLAUDE.md; procedures move to skills that load on demand. Sorting docs by rate of change is equally transferable — dated notes that expire, versus timeless architecture docs, versus live data fetched through MCP rather than pasted and left to rot.

Third, a caution: Faik's advice to let Claude maintain CLAUDE.md and never hand-edit it is the one piece I'd push back on. An agent's standing instructions are configuration, and unreviewed self-modifying configuration is how you get slow drift you can't explain three weeks later. Let the agent propose edits to its own map; review them like you'd review any config change. Also worth saying plainly: these workspaces accumulate strategy docs, user research, and meeting notes in plaintext folders people are encouraged to push to GitHub. Engineering has repo-hygiene reflexes for that. PM orgs mostly don't, yet.

Genuine shift, derivative packaging

The HN thread was lukewarm — 13 comments, several dismissing the post as AI-written, plus a familiar argument about whether PMs belong in the "how" at all. Fair enough on the packaging; there's now a cottage industry of near-identical Claude Code starter kits for PMs and designers, and this one's practices largely mirror Anthropic's own guidance.

But judge the direction, not the artifact. When engineers, PMs, and Anthropic's product line all independently land on files-over-prompts as the way agents accumulate competence, that's not hype — that's an ecosystem agreeing on an architecture. The filing cabinet, it turns out, was the moat all along.