Two Claude sessions, one plans/orchestrates and the other builds
retained text
I realise this is almost certainly old news or absolute simplicity to most of you but having looked at many posts, watched too many videos, and got drawn in too many times on TikToks I've eventually got a workflow working with Claude Code that I love. I just thought this might interest some that are like me, appreciate a clear distinction of roles but expect to be able to automate it without deep pain. I'm a simple vibe coder doing personal projects. Because of this I don't need anything sophisticated, but I was annoyed with using a single session and forgetting to change the model/effort, or using two sessions and having to copy and paste between them. If it matters, the project I'm working on at the moment is using Vite, React and TypeScript on a single Cloudflare Worker, with D1 (SQLite) as the database, all on the free tier, and a staging copy of the site. The code is in Github. I run two Claude Code sessions side by side in the Claude desktop app, and each has one job. **The planner (Opus)** * It lives in a folder outside the repo, with its own clone that always stays on `main`. It can also read the developer’s clone to see work in progress, but never changes it. * It plans but never codes. A hook blocks any file writes except for briefs and its own memory notes. npm, wrangler and PowerShell are denied, and git and gh are allowed only in a few exact forms, so it rarely stops to ask me for permission. * It turns my request into a written brief (a markdown file in `docs/briefs/`): why, the decisions already made and the reasons, scope and what’s out of scope, acceptance criteria, tests, and a “How to check it” section with screen widths and a JavaScript snippet that measures overflow. * For anything visual, it builds a clickable HTML mock-up and publishes it as a private Artifact, so I can look at it on my phone before any code is written. **The developer (Haiku or Sonnet)** * It works in the real repo. It reads the brief, branches, builds, runs the checks (typecheck, about 570 tests, lint), opens a PR and pushes the branch to staging. * Questions about the brief go to the planner, not to me. The planner answers what the brief or the code already answers and only passes on decisions that are really mine. **How a sprint runs** 1. I describe what I want. The planner reads the code, asks me anything it can’t work out, and says back in a few lines what will and won’t be built. 2. I say yes. It writes and pushes the brief, then sends it to the developer session itself with a cross-session message, so there’s nothing for me to copy and paste. 3. The developer builds it without interaction from me, opens the PR and messages the planner with the PR number. 4. The planner waits until the developer is finished and then automatically reviews the diff and CI, then opens the dev server in the desktop app’s built-in browser. It checks the screens at phone and desktop widths and measures overflow with JavaScript rather than trusting screenshots. 5. If I want to start planning the next phase I can while the development is going on. The Planner still maintains the checks on the development session and on we go. 6. It gives me one summary: what was built, what it checked, what it couldn’t check, and my steps. 7. I test on staging and say “go”. The planner merges, and Cloudflare deploys the live site. **Things that made it work** * **A status band.** It’s a small custom plugin (a “mod”) that shows the sprint’s stage above the prompt (questions, mock-up, briefed, building, reviewing, ready), plus what the developer session is doing right now. It also wakes the planner when the developer stops, because the built-in “tell me when that session goes idle” option doesn’t work between desktop sessions. * **Clear lines on what stays mine.** Database migrations on the live database, merging, and any change to live data are always my call, whoever asks. * **The project’s** [**CLAUDE.md**](http://CLAUDE.md) **is the rulebook.** It’s long, and most of its rules exist because something went wrong once. Briefs point to it instead of repeating it. * **The planner’s memory is mirrored into the repo**, so the setup moves between machines.
document fetched 2026-10-10 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1x1nhq6/two_claude_sessions_one_plansorchestrates_and_the