Running an AI Team in Claude

The boss, the chief advisor, the coordination hub, the members.

Running an AI Team in Claude, slide 1: four levels, three model tiers, one line nobody crosses; what you do defines the role, what you don't do defines the boundary. Running an AI Team in Claude, slide 2: I Structure · Four levels, one line nobody crosses. Boss, chief advisor, tech lead, members; the chief advisor sits on the review chain, not the command chain, and the line nobody crosses runs between the boss and the members. Running an AI Team in Claude, slide 3: I Structure · What you do defines the role; what you don't do defines the boundary. Both columns get written: prohibitions only remove behavior and a pile of don'ts does not add up to a person. Roles are written in positive instructions, boundaries guarded by negative ones. Running an AI Team in Claude, slide 4: II Model choice · Model tiers: the tier sets the default, temperament can override. Final-review, execution and routine tiers; the tier is a default, not a ranking, and within a tier you pick by the temperament of the task, not by version number. Running an AI Team in Claude, slide 5: II Model choice · What separates the tech lead from the chief advisor is not rank but threshold. Three-stage filtering, each stage handling only what the previous let through; a threshold set too low costs more than one set too high, because every unnecessary escalation spends the team's scarcest tier of judgment. Running an AI Team in Claude, slide 6: III Dispatching · What a work order must contain. Open with the plan version number; spell out the deliverable and acceptance criteria; after dispatching, check the other side is still running; on receiving a list of constraints, ask whether they are enough taken together. Running an AI Team in Claude, slide 7: III Dispatching · Dispatching is a closed loop, not a single delivery. A one-way action has no receipt: the sender thinks it arrived, the receiver never got it; the blind spot only disappears when the receipt becomes part of the loop. Running an AI Team in Claude, slide 8: III Dispatching · How Agents talk to each other: three rules learned the hard way. Address by session ID, not by name; read the return value, only the sent line counts; send important reports twice. Running an AI Team in Claude, slide 9: IV Operations · Reporting and progress: how the boss actually asks. Check progress, balance the load, find the idle, subscribe to completion; at this level it is almost no different from running a human team. The real difference: a person will say they are stuck, a session won't. Running an AI Team in Claude, slide 10: IV Operations · The two sides of the escalation threshold, and handing down what you know. Too low takes up rulings, too high means acting on your own; a list of pending decisions rots like a to-do list, only more expensively. Handing down knowledge is an action, not a hope. Running an AI Team in Claude, slide 11: IV Operations · One-way actions have to be fixed on both ends; fix one and it still leaks. Estimates must be labeled as estimates and differences reported after measuring; do only the left column and errors go uncorrected, do only the right and the receiver is cleaning up after the sender. Running an AI Team in Claude, slide 12: IV Operations · When writing it down, keep three kinds of record apart: debt, design, conditional pass. Writing design as an exemption leaves a fake to-do list; writing debt as design writes the debt off. Running an AI Team in Claude, slide 13: V Constitution · Four layers of text, four kinds of edit rights. Who can change it decides whether it is law more than what it says; nobody edits the charter, the boss has sealed it, and the archive only takes in. Running an AI Team in Claude, slide 14: V Constitution · How a session with no memory knows who it is. Every role reads four documents, members three more; start work means load first, and a rule that was not loaded does not exist. Running an AI Team in Claude, slide 15: V Constitution · Written down is not the same as findable. If it can be caught in code, don't only write it in a document; write keywords the way they will be searched; make the title name the mistake it prevents. Running an AI Team in Claude, slide 16: VI Secondment · Working with an outside team: the temporary crew. Send a fork, not a newcomer; dual reporting with clear responsibilities; a temporary crew still writes its constitution first. The shorter-lived the crew, the more the text is its only continuity.