How to Make Claude Code Collaborate — Doorae and Conductoring

If you've pushed Claude Code hard, you know this complaint. You spin up subagents in parallel, and they quit halfway. Or worse — they finish, and they've built the wrong thing.

Here's why it happens: a subagent has no way to ask.

Look at the structure. An agent hands down an order; the subagent takes it and runs. One direction, one time. But real work always hits a fork — A or B? Pile it here or defer it? A human worker would turn around and ask. A subagent has no channel. So it does one of two things: plows ahead on a guess (an accident), or stops dead (a runaway). Either way, you end up chasing behind it, re-reading everything it made, cleaning up. That's where your time and your tokens go.

Start from an honest premise: a perfect work order does not exist. The moment a worker measures the actual code against the instructions, something missing shows up. It always does — human organizations have known this for centuries. So the problem isn't "write better orders." The problem is: when the gap appears, who answers?

Should it be you? Every fork, every "A or B?", landing in your lap?

Think of an orchestra. If the musicians stopped mid-performance to ask the theater owner "should I hit this note harder?", that's not an orchestra — that's chaos. That's what a conductor is for.

So here's how myDoo runs Claude Code — we call it Conductoring.

Always stand up a PM first. You talk with the smartest, most expensive model — Fable — about the whole picture: the logic, the goal, what "done" looks like. That conversation is the delegation. Get it right, and everything downstream stands on it.

The PM turns around and calls up coworkers. Each one gets a written work order. And the PM picks the right model for each job — Opus for the hard parts, Sonnet for the bulk. The expensive model thinks; the cheaper models type. This is also, quietly, how you save a lot of tokens.

Notice the word: coworker, not subagent. Co-, not sub-. These aren't hired hands you fire-and-forget — they're on the ledger by name, they write reports, they exchange letters, they show up on the board.

A coworker measures reality. When it finds the gap in its work order, it writes a question — and stops. It does not build past an open question. This is the discipline that changes everything: a coworker that stops is nothing like a subagent that bails. Stopping is protocol. Bailing is an accident.

The PM answers — within the intent you delegated. The coworker picks the answer up and keeps going. Questions get absorbed at the PM level; only the calls that genuinely need your judgment reach you.

And it scales. Stand up multiple PMs — one per desk, per domain. PMs talk to each other, trading letters to sort out where their territories overlap. Smart PMs, disciplined coworkers, and you talking to exactly one layer.

We named this exchange Doorae — Request · Answer · Exchange. It's an old Korean word: a doorae was a village work crew, neighbors pooling their labor in each other's fields. Others call this "orchestration." We think the village had it first.

And the letters aren't just records. When a coworker writes one, the recipient's doorbell rings — "a letter for you, read the board." The message gets delivered; the file remains. That's the trick: conversation that leaves a paper trail. Our echo test clocked a full round trip — question out, answer back, signed and filed — in 31 seconds.

You watch it all on the Work Board: every desk, every PM with its tree of coworkers underneath — green pulse for working, a resting seal for an idle PM, and a folded stack of "done today." You're not the switchboard operator anymore. You watch the letters flow, and you answer only what needs an owner.

One more thing, and it might be the best part. Every step of this — the design brief, the work orders, the questions, the answers, the completion reports — is a document. Which means the whole process becomes an asset: what you tried to build, where the plan was thin, what the workers discovered. It's all there, readable, reusable for the next job. Building a working AI organization from scratch, we ended up rediscovering something very old: paperwork. Work orders, memos, sign-offs. It turns out administration wasn't bureaucracy — it was how organizations survive members who sleep, die, and get replaced.

Claude Code doesn't quit because it's dumb. It quits because it has no one to ask.

Give it a conductor.