Building with agents, written down.
How we plan, run, and review work done by coding agents, on our own projects. One job per post. No roundups.
Scope an existing repo: planning with AI when the code is already there
Planning tools assume a blank page. Your project isn't one. How to get a plan that knows your stack, your conventions, and what already exists, without writing a single file into the repo.
4 min readPlan a feature before any agent touches code
The expensive part of a feature isn't writing it. It's finding out, halfway through, what it actually was. Ten minutes of thinking first is the cheapest thing you'll do all week.
3 min readHow to fix bugs with an AI agent without it breaking something else
The agent fixes the bug and quietly breaks two other things. That isn't a model problem. It's a task problem. The shape of a bug task an agent can finish cleanly, and the loop that catches it when it can't.
3 min readCLAUDE.md vs a living plan: what each one is for
You need both. A CLAUDE.md holds the rules. A plan holds the state. Trouble starts when a rules file is asked to hold the state, and it starts quietly.
3 min readHow to run Claude Code in the cloud: every option compared
A VPS, a hosted workspace, or a run driven by the plan. Every way to run Claude Code somewhere other than your laptop, what each costs you in setup, and how to get a pull request back without babysitting a terminal.
3 min readRun Claude Code from your phone: queue a task, come back to a PR
No SSH apps, no terminal on a six-inch screen. Queue a scoped task to a cloud run from your phone, answer the agent's question from a notification, and review the pull request when you're back at a real screen.
2 min readClaude Code keeps losing context. Give it a plan instead of a paragraph.
Every new session starts from zero. CLAUDE.md helps until it sprawls. The fix is a live plan your agent reads over MCP, the same one in every session, every editor, and every teammate's chat.
2 min read