Project 04 — Your First Cowork Workspace

Enforces: delegating and supervising a real multi-step task in a Cowork workspace (from introduction-to-claude-cowork) Surface: Claude Cowork · Time: ~50 min · Difficulty: 🟡 core

Why this project

Chat is a conversation; Cowork is a workspace where Claude does multi-step work over files and tools while you supervise. The skill the course is really teaching is supervision — handing off a task, watching each step, and correcting course before a small wrong turn compounds. You only build that muscle by running one real task end to end.

What you'll build

A Cowork workspace that completes one genuine multi-step task (research → draft, or analyze → summarize), plus a short written reflection comparing Cowork to plain chat and to Claude Code.

Steps

Set up the workspace — open Claude Cowork and create a workspace for one concrete deliverable. Give it a clear name and, if the surface supports it, drop in any reference files the task needs (a brief, a spreadsheet, source notes).
Pick a genuinely multi-step task — something that needs sequencing, not a one-shot answer. Good candidates:
- Research + draft: research a topic across a few sources, then draft a one-page brief.
- Analyze + summarize: load a data file, find the three biggest trends, write them up for a non-technical reader.
Write the delegation prompt — state the goal, the constraints, the output format, and explicitly ask Claude to work step by step and check in before big moves. This is the Description skill from p03-ai-fluency-delegation-audit applied to an agentic surface.
Supervise each step — don't fire and forget. After each step, do one of:
- Approve — the step is right; let it continue.
- Correct — it drifted; say specifically what's wrong and what to do instead.
- Redirect — the approach is wrong; change the plan. Keep a running note of every correction you made — those notes are the evidence you supervised rather than rubber-stamped.
Verify the deliverable — apply the Diligence checklist: facts sourced, claims checked, output actually usable. Fix what's left yourself.
Reflect — Cowork vs chat vs Claude Code — write a short comparison: when would you reach for each? Rough guide to react to:
- Chat — single questions, quick drafts, thinking out loud.
- Cowork — multi-step knowledge work over files, where you supervise.
- Claude Code — work in a codebase, where the deliverable is code and the loop is build/test/run.

Acceptance criteria — you're done when

You created a Cowork workspace scoped to one real deliverable.
You delegated a genuinely multi-step task (not a one-shot question).
You supervised each step and logged at least two concrete corrections or approvals.
You verified the final deliverable against a Diligence checklist and fixed remaining gaps.
You wrote a Cowork vs chat vs Claude Code comparison with a clear "reach for X when…" for each.
You journaled what surprised you in learning-journal-template.

Stretch goals

Re-run the same task in plain chat and note exactly what was harder or worse without the workspace.
Deliberately let one wrong step slide, then watch how far the error propagates before you catch it.
Identify one task in your real week that should move to Cowork and one that should stay in chat.

Self-assessment rubric

Level	Signal
🟢 Got it	You supervise actively, correct early, and can say which surface fits which job.
🟡 Almost	You delegate well but tend to check the result only at the end, not each step.
🔴 Revisit	The workspace felt like chat with extra steps; re-watch introduction-to-claude-cowork.