Skip the todo – just write the prompt
Plus: how I orchestrate Claude Code with Zo Computer
These days, Claude Code is being discussed a lot. It feels like a big new wave of people tried Claude Code over the holidays, and are discovering how useful coding agents can be. It’s amazing to see non-technical folks getting this excited about a CLI – a tool where the only interface is text.
When it comes to AI coding, I’ve never been too attached to any specific tool, because we’re living through the great industrialization of code. Generating code is a commodity, and the tools for doing it are fungible.
Personally, I’m all over the place. I’m in Cursor. I’m in Zed with Claude Code and ACP. I’m texting my Zo to make a PR with Claude Code while I’m walking to the beach. I’m firing off tasks in Cursor Web from my iPad, just to feel something. The bottleneck as a builder is increasingly just the limits of my human brain. Ever since I started working with AI in 2023, I’ve been obsessed with overcoming those limitations by experimenting endlessly.
Recently, the big shift for me has been realizing I should skip logging the todo, and instead just write the prompt. Before, I’d file small todos for myself in Linear. Now, simple tasks go straight to PR. Even complex tasks can be kicked off with a prompt to write a plan (see this post for more on my planning technique).
Orchestrating Claude Code with Zo
The latest addition to my workflow has been using Zo to orchestrate Claude Code agents. Zo can, of course, do lots of coding (and non-coding) tasks all by itself. But using Zo as the orchestrator for various coding tools (running as “headless” CLIs) feels like the next-level unlock. (Here’s my Zo x Claude Code prompt, and an example chat).
This setup has all sorts of benefits for my day-to-day work. These benefits are totally possible to achieve in numerous other ways, but Zo Computer makes it easy.
I love that I can:
Fire off multiple small ideas in parallel, and trust Zo to create isolated branches and PRs, using git worktrees.
Work from anywhere. I can text my Zo with a prompt, and review the PR from the
GitHub mobile app.
Ask my Zo to spin up a service and send me a preview URL to verify changes.
Connect to a Claude Code session in Zo using Termius from my phone, or my iPad.
Connect to my Zo workspace using Cursor over SSH.
Ask Zo to look at our latest SEO research and suggest updates to our CMS and marketing site. I have a Zo agent chugging along on SEO research, searching in Ahrefs and writing to a SQLite database in a folder on my Zo.
Ask Zo to summarize last week’s activity in our codebase and write a report file, in another folder on my Zo.
The list goes on.
Whether you use Zo or not, I highly recommend creating a setup for yourself like this. Even if you don’t do much coding, it’s worth experimenting with this type of workflow for non-coding tasks. AI coding agents are really just “working with files” agents – they can do a lot more than coding projects. But when you take these agents beyond coding, you need a different interface. That’s Zo.
This slice of my workflow is just a taste of what’s possible in Zo Computer – where we’re on a mission to build the perfect AI tool: for ourselves (a bunch of computer nerds), our friends & family, and maybe you, too :)
This resonates so much with how I've been thinking about AI workflows lately. The shift from "manage todos" to "just prompt it" is subtle but genuinely changes how you approach work. It's less about tracking what needs to be done and more about describing what you want to exist.
I've been building something similar with Claude Code - an AI agent I call Wiz that acts as my universal entry point for everything. The key insight I landed on was giving it persistent memory and the ability to spawn sub-agents for different domains. So instead of switching between tools or remembering which system handles what, I just talk to Wiz and it figures out the routing. "What should I post on TikTok?" goes to the content agent. "How's the influencer project?" pulls up that project's full context.
What you said about working from your phone really clicked too. There's something freeing about treating your development environment as something you talk to rather than something you sit at. The mental model shifts from "I need to be at my computer to code" to "I can describe what I want from anywhere."
The parallel agent execution is where things get interesting. I've found that having agents maintain their own memory files - what worked, what didn't, domain-specific preferences - means they actually get better at their jobs over time. It's like building institutional knowledge into the system itself.
I wrote up the whole architecture and philosophy behind Wiz if you're curious about another take on this approach: https://thoughts.jock.pl/p/wiz-personal-ai-agent-claude-code-2026