Cowork for Teachers: Handing Off Work That Carries Forward
Claude for Teachers includes Cowork and Claude Code, so Claude can run multi-step work autonomously. A practical guide to what a teacher should hand off, what to keep, and how to delegate work that carries forward without you.
Most people meet AI as a chatbot: you ask, it answers, the exchange ends, and nothing happens until you ask again. That model is fine for a quick question and useless for the actual shape of a teacher's work, which is rarely one question. It's a chain — read these diagnostics, then draft plans against them, then differentiate each plan, then package the materials. A chatbot makes you the glue between every step, holding the whole chain in your head and clicking it forward link by link. The bottleneck was never the model's intelligence. It was you, doing the carrying.
Claude for Teachers changes what you're working with. It includes Cowork and Claude Code, which means Claude can carry multi-step work forward autonomously — not answer one prompt and stop, but run a job through its stages while you do something else entirely, like teach. That's the difference between a tool you operate and work you delegate. This piece is about the judgment that difference demands: what to hand off, what to keep, and how to give an agent enough to run without giving away the parts that have to stay yours.
Chatbot versus agent, and why it matters for teachers
The distinction sounds abstract until you feel it in a real task. Suppose you want next week's five lessons differentiated across three readiness levels, informed by last week's diagnostics. With a chatbot, that's a dozen prompts, each one waiting on your attention, each output pasted into the next request. You're present for every step, which means the task takes as long as your available attention — and your attention is the scarcest thing you own.
With an agent, you describe the whole job once and hand it over. Claude reads the diagnostics, drafts the plans against your standards, produces the tiered versions, and renders the materials — carrying its own output forward from one stage to the next without waiting for you to shuttle it along. You come back to a finished draft to revise. The work didn't get smaller; it got lifted off you. For a profession defined by having thirty things happening at once, being able to set a multi-step job running and walk into fourth period is not a convenience. It's a structural change in what one teacher can get done in a day.
A chatbot answers when you ask. An agent does the job while you're doing something else. For a teacher, whose attention is the binding constraint, that's the whole game — the win isn't a smarter answer, it's not having to be present for every step.
What to hand off
The skill worth building is knowing which work is safe and valuable to delegate. The best candidates share a shape: mechanical, multi-step, well-specified, and reversible — where the output is a draft you'll review, not a decision that ships unchecked.
| Good to hand off | Keep in your hands |
|---|---|
| Synthesizing diagnostics and rosters into a class picture | Deciding what an individual student actually needs |
| Drafting standards-aligned lesson plans to revise | Approving the plan you'll actually teach |
| Producing tiered differentiation across readiness levels | Judging which student gets which version |
| Reviewing exit tickets and drafting tomorrow's adjustment | Reading the room and the relationships in it |
| Rendering materials into printable documents | The teaching itself — the human part |
The pattern in the left column is that each is a first draft of something, and drafts are exactly what agents are good at. The Claude for Teachers workflows are built to produce artifacts you revise, not verdicts you rubber-stamp — the lesson-planning and differentiation workflows both hand you materials to sharpen. When work is reversible and reviewable, delegating it costs you nothing and buys you back hours. The 4pm exit-ticket review is the purest version: a job that runs on its own schedule and simply has a draft waiting for you.
What to keep
The right column matters just as much, and getting it wrong is the failure mode to avoid. Some work should never be handed off, not because Claude can't attempt it but because it's the part of teaching that only a human, in the room, should own.
You keep the judgment about individual students — what this particular kid needs, why she's disengaged, whether his diagnostic gap is a skill problem or a confidence problem. Data informs that call; it doesn't make it. You keep final approval on anything students will actually see or be graded on. And you keep the relationships, which is the point Randi Weingarten, president of the American Federation of Teachers, made about the whole design: the goal is to "give them more time for the human relationships at the heart of learning." The reason to hand off the mechanical work is to protect the human work, not to automate it. An agent that drafts your lessons so you can spend fourth period actually seeing your students is doing exactly what it's for. An agent making the calls that require knowing those students by name is not.
This is also where the responsible-use guidance the series keeps returning to lives: the AI Fluency for K-12 Teachers course, co-created with Teach for America, is built around precisely this question of which classroom tasks AI is suited for. Delegation without that judgment is just abdication with extra steps. For teachers who want to explore agent-style teaching tooling beyond Claude for Teachers, the marketplace education agent skills collection is a good place to start.
How to delegate well
Handing off well is a learnable practice, and it looks a lot like briefing a capable student teacher. Be explicit about the goal, the constraints, and what "done" means — the standard you're on, the class you're planning for, the format you want back. Give the agent the context it needs to run without you: the diagnostics, the roster, the curriculum you're following. Vague instructions produce vague work you have to redo, which defeats the entire purpose of handing it off.
Then let it run, and review the output as a draft, not a verdict. The first few times you delegate a given job, check it closely and correct what's off; those corrections teach the agent your standards, and the next handoff comes back closer to right. Over a few weeks you build a set of jobs you trust it to carry — and a clear line around the ones you don't. Because the work happens through Cowork and Claude Code, and Claude for Teachers holds your data under its own FERPA-aligned teacher terms with nothing you share used for training, delegating class-related work is designed to be safe to do, not a risk you're quietly taking.
The teacher as director
Step back and the shift is bigger than any single workflow. When Claude can carry multi-step work forward on its own, your role moves from doing every step to directing the work — deciding what needs doing, handing off the parts a well-briefed agent can run, keeping the parts that need your judgment and your presence, and reviewing what comes back. That's not a smaller job. It's a higher-leverage one. The same teacher, with the mechanical carrying lifted away, can differentiate for the whole room, act on their data every week, and still have the attention left over for the students in front of them.
That's the destination the Claude for Teachers series has been building toward: not AI that teaches for you, but an agent that carries the work so you can spend more of yourself on the part only you can do. To see how this autonomy shows up across a teaching week, read Claude Code + Cowork in the Classroom. Verification is free through June 30, 2027 — and the first thing to hand off is whichever multi-step job you're most tired of carrying yourself.
Part of the Claude for Teachers series. Related: The 4pm Exit-Ticket Review That Runs Itself · Claude Code + Cowork in the Classroom. Browse AI tutoring skills or more builder insights.