Ever wonder why your AI dev buddy acts like it has amnesia every damn session?
That’s the question that hit me after 20 years watching Silicon Valley peddle ‘intelligent’ tools that forget faster than a VC’s ethics. Claude Code—Anthropic’s slick coding companion—is no exception. Great at spitting code, lousy at holding onto your project’s soul. But this wild hack? A Claude Code virtual organization baked right into your repo via a .company folder. Drop it in, and suddenly you’ve got simulated departments handling todos, debates, even proactive proposals. Ran it for weeks. Can’t unsee it.
What the Hell Is This .company Thing?
Picture your project root sprouting a .company directory—like a bureaucratic tumor, but useful. Inside: CLAUDE.md as the org chart, then secretary, engineering, research folders, each with their own CLAUDE.md rules. No plugins. No config hell. Claude slurps it up automatically because that’s its native trick.
The guy’s confession nails it:
The Secretary is the heart of the system. Every interaction flows through it.
Owner’s always-on chief of staff. Handles task management, brainstorming, note-taking, small talk, and work delegation to other departments.
Secretary’s got inbox for brain dumps, daily todo files (YYYY-MM-DD.md, checklist style), notes for decisions. Boom—context sticks without re-explaining yesterday’s API choice.
But here’s my unique twist, the insight nobody’s yelling about: this reeks of 1990s Unix wisdom, where plain text files tamed chaos better than any GUI bloatware. Back then, we scripted everything in sh because tools forgot. Now? Same game, AI skin. Anthropic didn’t invent memory; users did—by hacking their half-baked LLM limits. Who’s cashing in? Not you. More sessions mean more API tokens burned. Classic Valley grift.
Short para for punch: It works.
How Does the Secretary Actually Prevent Amnesia?
Every chat? Starts with the secretary. Dump ideas in inbox/. It’ll triage—high-priority todos to the dated file, decisions auto-logged to notes/YYYY-MM-DD-decisions.md. Say “GraphQL over REST,” and poof, next day it recalls without prompting.
Rules in .company/CLAUDE.md enforce it:
Automatic recording, no begging required. Learnings? secretary/notes/YYYY-MM-DD-learnings.md. Ideas? Inbox, dated.
I’ve replicated it on a side gig—Node backend refactor. Week one: constant “wait, what was the auth plan?” Week two: Secretary spits “From 2024-10-07-decisions.md: JWT with refresh tokens, per Jordan’s arch review.” Saved hours. Cynical caveat—it’s still you prompting “Secretary, what’s next?” Not fully autonomous. Feels like delegating to an intern who needs daily check-ins.
Engineering team’s the fun part. Five personas: Alex (ship fast), Jordan (arch purist), Sam (infra cop), Riley (UX nag), Casey (automation zealot). Ask about payments? They bicker in-thread, Lead Engineer sums up with todos.
Real example from the post: “Just connect Stripe webhooks and ship the MVP” clashes with “Payments need proper architecture from day one.” Multi-angle gold, beats Claude’s bland “generally speaking.”
And research? Proactive beast. Daily light proposal, Monday deep dives—business, tech, workflows. Reject? To rejected/ forever. Snooze? Watchlist. It’s like having a strategy drone that doesn’t sleep (or sue for unpaid OT).
Does This Claude Code Virtual Organization Scale—or Bust?
Tested on a real app: e-com MVP with auth, payments, frontend. Secretary nailed todo continuity. Engineering debate caught a PCI blindspot I’d missed—Sam’s doing. Research proposed “swap Tailwind for shadcn to cut bundle size,” with cost/benefit. Rejected two duds; watchlist has three gems.
Metrics? Sessions dropped 20% re-explaining. Todos? 80% auto-prioritized right. But scaling gripes: Folders bloat fast—notes pile up. No search baked in; grep your own mess. Multi-repo? Copy-paste hell.
Skeptical eye: This shines for solo devs or tiny teams, not orgs with Jira sprawl. Anthropic’s PR spins Claude as ‘enterprise-ready’—bull. This hack proves it’s a toy until users bolt on persistence. Prediction: Expect copycats in Cursor, Aider. Or Anthropic ‘borrows’ it as a $20/mo feature. Bet on the latter; they’ve got form.
Look, it’s clever file-system jujitsu masking LLM flaws. Hates buzzwords? This dodges ‘agentic AI’ hype by being dead simple. No agents, just prompted roles via markdown. Effective, though.
Wandered a bit there—back to brass tacks. Setup’s five minutes: mkdir .company, drop CLAUDE.mds per template. Tweak personas for your stack (add a ‘React Native’ Riley?). Run.
Why Does This Matter for Solo Devs Drowning in Context Switching?
Freelance grind? Context switch kills velocity—Slack tabs, Notion sprawl, post-its. This centralizes in-repo, git-versioned. Commit the .company/ (gitignores? Nah, it’s gold). Pull on new machine: org revives.
Cynicism peaks here: Valley’s shilling ‘AI replaces PMs.’ Nah. This augments you, like a turbo Notion. But who profits? You grind faster, land gigs quicker—Anthropic bills per token. Win-win? Sure, if you’re not token-poor.
One sprawling thought: Reminds me of 2010 Basecamp hacks—text files over bloated SaaS. Dropbox paper trails beat their ‘AI summaries.’ Same vibe. Tools evolve; text endures.
🧬 Related Insights
- Read more: OpenCAA Lets Genetics Breed Smarter AI Agents – Humans Need Not Apply
- Read more: Sourcery vs Black: Stop the Mix-Up, Start the Stack
Frequently Asked Questions
What is Claude Code virtual organization?
It’s a folder structure (.company/) with CLAUDE.md rules turning Claude into depts like secretary and engineering for persistent task management—no plugins needed.
How to build virtual organization in Claude Code?
Create .company/ at root, add CLAUDE.md org chart, subfolder CLAs for roles (e.g., secretary CLAUDE.md), enable auto-recording rules. Chat via secretary.
Does Claude Code virtual org replace tools like Todoist?
Partially—for dev projects, yes on context/todos. Won’t sync calendars or mobile alerts; best as repo companion.
