Coworking with Claude

It's Monday morning. Before I open Figma, before I open VS Code, before I check Slack, I open Notion. A fresh weekly plan is waiting — written at 7am by an agent that spent the night reading my Slack threads, scanning new Pulumi blog posts, parsing what's current in the infrastructure-as-code industry, triaging open customer tickets, and combing my email. Its output: five to eight actionable product design improvements, ranked, with rationale, and a draft presentation in a state I can share with product and engineering after a fifteen-minute read-through.

This is what most of my week looks like now.

I'm calling the practice coworking with Claude. Not because Claude does the work for me — it doesn't. Because Claude is always on, always available, and increasingly the first thing I turn to when context is the bottleneck and effort isn't. The shape emerged over the past few months, after I saw Jenny Wen write about treating an agent as part of her team. I lifted the spirit of her workflow and made it mine.

The Principal Product Designer skill

The first piece is a persona. I used ChatGPT to author a Principal Product Designer skill — a long-form document that captures how I want every agent task to be framed. The skill describes the level of work I'm trying to do (Principal-track product design at Pulumi, AI-native, design + code, design systems, technical user audiences), the things I care about (judgment over volume, decisions over outputs, accessibility as default), and the things I don't want (sycophancy, executive jargon, premature closure).

I attach this skill to every conversation that needs context. The agent reads it before it does anything else.

That sounds like a small move. It isn't. You design at the level you ask your agent to design at. A vague prompt yields vague work. A persona document framed at Principal-track altitude produces output I can actually use, instead of output I have to fix.

Always aim for the next level. The skill is how I encode that into the substrate.

Four scheduled tasks

The Principal skill runs ahead of every task. Underneath that, I run four scheduled tasks that handle the recurring shape of my week.

Monday — weekly plan. Every Monday at 7am, an agent generates a fresh Notion entry plus a draft presentation. Its inputs: my Slack threads from the previous week, IaC industry news, open customer tickets, Pulumi blog posts, and my email. Its outputs: five to eight actionable product design improvements, ranked, with rationale. The week starts with a frame instead of a flood.

Daily — currents. Every morning, a smaller task runs to keep me current on the state of the art across AI-native design, infrastructure-as-code, and developer experience. The agent surfaces and summarizes — research papers, product launches, design-system releases, sharp essays. One digest, a handful of items, nothing more.

Hourly — Slack pulse. Every hour during working time, an agent scans Slack for at-mentions and unresolved threads I'm in. For each, it suggests how to reply — tone, content, a draft if the situation warrants it. The point isn't to automate the replies. The point is that I'm never surprised. By the time I open Slack, I already know the shape of what's there. That alone keeps me grounded and calm.

Friday — weekly summary. Every Friday afternoon, a final agent task produces a summary — what shipped, what slipped, what's in flux, what to carry into next Monday's plan. It closes the loop. It also makes Monday's plan warmer, because the agent has working memory across the seam.

What changes

A few things change when you set this up:

• The week starts with context, not clutter. Monday's first hour used to be inbox triage and Slack catch-up. Now it's a focused read-through of a plan the agent prepared overnight.
• Work shifts asynchronously across time zones. The agent runs while I sleep. I wake up to outputs I didn't have to grind for. My hours go to the design problems, not the context-gathering tax around them.
• I respond from preparation, not reaction. The Slack pulse, in particular, has changed the cadence of my day. I no longer dread opening Slack after a meeting and discovering three threads I should've been in. I already know about them. I have a draft. I respond at human speed instead of catch-up speed.
• I notice when the agent is wrong. This is the most important part. A meaningful fraction of what each task produces is mis-prioritized, context-deaf, or just wrong. The part that's right multiplies my throughput. The part that's wrong forces me to engage with what matters. That's the whole point of an AI-native practice: judgment over output.

Where this came from

Jenny Wen was thinking about this before I was, and her writing pulled the shape together for me. The four tasks aren't hers exactly, but the idea — that the agent is part of the team, that you commit to scheduling its work and reading its outputs the way you'd commit to a teammate's — that came from her.

Coworking, not delegating

I want to be careful with the framing. I do not delegate to Claude. I cowork with it. The distinction matters. Delegation implies handing off and reviewing. Coworking implies sitting next to and steering. The Principal skill is how I steer. The four scheduled tasks are how I keep context warm. The work is still mine. The judgment is still mine.

What's different is that I show up to it prepared.

— Dylan