The Agent Workspace Starter Kit is a free, two-repo setup for running more than one project with an AI coding agent: an executive-assistant repo that knows you, and a product-copilot repo per product that knows the product. It comes as a guide, two paste-in prompts that scaffold both repos in about an hour, and the operating habits that keep them useful for months. We call the idea Builder OS. Everything at NocodePro runs on it, every day. It works with Claude Code, Codex, Gemini CLI, Qwen Code or any agent that reads a brain file and a skills folder.
The problem it solves
Most people run their agent as one chat in one folder. It works, until the projects multiply. Then every session starts with you re-explaining who you are, what you're building and how you like to work. The brain file (CLAUDE.md in Claude Code, AGENTS.md in Codex) turns into 400 lines of "if working on X do Y", and the agent knows either your strategy or your product code, never both well.
The fix is not a longer brain file. It's two of them, with different jobs.
How it works: two repos, two jobs
The executive assistant knows you: goals, priorities, team, preferences, the decisions you made and the mistakes you don't want repeated. You plan here: the start of the week, a priority shift, a new project, a reminder for a date. About 10% of your time.
The product copilot, one per product, knows the product: database schema, design system, user flows, API contracts, content and voice. You build here. The other 90%.
Each has its own brain file, its own decisions log, its own references. They don't share files. You carry context between them, and two small conventions, a project index in the assistant and a decisions log in each repo, keep them in step. The guide also covers the nested variant, where product repos live as their own checkouts inside the assistant's projects folder.
What's inside the kit
Ten files, 35 KB, no dependencies beyond the tools you already build with:
- Getting-started guide. Why two repos, a mapping table for every agent, the setup step by step, the operating habits, the day-to-day workflow, troubleshooting, and a page on where the setup goes next.
- Executive assistant init prompt. Paste it into a fresh session. Four guided phases: folder structure, tools and rules, a six-section onboarding interview, then the agent writes your context files and brain file from your answers.
- Product copilot init prompt. The same for a product. Product questions only.
- Session handoff template. The note one session writes for the next, so nothing restarts from zero.
- Global rules starter. The working discipline that applies in every repo: how to handle ambiguity, verification before claiming done, how a work block closes, git conventions, the scheduled-tasks rule. Every line was broken first.
- Two hooks for Claude Code, with a one-line fallback for other agents: scheduled tasks that surface at session start, and a guard that keeps
.envfiles out of commands and commits. - Changelog, and a README that says what to read first.
The operating habits
The two repos are the structure. Eight habits are what made the structure hold across five products for five months:
- The two logs. A decisions log with the reasoning, and a mistake log the agent reads before attempting an approach. Entries link to earlier entries of the same kind. The families are the real lesson.
- Curated memory. One fact per note, typed (about you, feedback, project, reference), with the why attached. Logs-as-memory for agents without a memory feature.
- Global rules vs project rules. If a rule would be true in a repo you haven't created yet, it's global. That's how the project file stays under 150 lines.
- Scheduled tasks. A dated file and a session-start hook. "Remind me on the 25th" means something.
- Session handoffs. A note from the closing session to the next one. First message: "Read this file and continue from it."
- Handing a repo to a co-founder. A bundle they unzip and open in their agent. The agent does the install; the human reads one handbook. We've done it twice, about an hour each.
- The tool-integrations table. Three columns in the brain file: what's connected, how, for what. The agent stops asking.
- Closing a session. A TL;DR with owners, and the context files kept current.
Which agents
Built and run daily on Claude Code, written so it carries over. The guide's mapping table covers brain-file names, global rules files, context imports, skills folders, hooks and memory for Codex, Gemini CLI and Qwen Code, with plain-file fallbacks where an agent has no hooks or memory. Both prompts open with a paragraph that tells a non-Claude agent how to translate every name, so the same file runs anywhere. Claude Code and Codex paths are the ones we verify in daily use. For Gemini CLI and Qwen Code, check the current docs for the skills folder.
Setup time
- Executive assistant: about 30 minutes, most of it the interview.
- Product copilot: about 30 minutes.
- Operating habits: 15 minutes to read, a week to feel.
After that, every session starts warm.
Who it's for, and who it isn't
For builders and founders running more than one project with an agent as their daily driver, and solo builders starting repo number one who don't want to rebuild later.
Not for a one-off script or a weekend project. Two repos are overhead you feel on day one and repay in week two. If there is no week two, one folder is fine.
What changed in version 2
The first version (April 2026) was two folders and one prompt, written for Claude Code to onboard a co-founder. Version 2 (September 2026) renamed it, made it agent-agnostic, added the second prompt, the handoff template, the global rules starter and the hooks, and wrote down the eight habits. The habits are the product. The folders are the container.