Odd Bits / Builder kits
Team-in-a-box
When you split work across several Claude Code sessions, the hard part is not the prompts. It is who owns what, how they hand work over, and what happens when one of them is confidently wrong. This is a clean, generic version of the setup I use for that, written from scratch so you can read all of it in an hour and change it to fit you.
- What you get
- Roles, a shared board and handoff, review and QA templates
- Gate rules and the lessons behind them
- A scrub check script for your own copy
- Works with
- Claude CodeMarkdown
- Version
- v1.0, updated 2026-10-19
Email for an invoice: $9.99Paid by invoice for now. Files arrive by email once it clears.
How you get it: email me. I send a USD invoice through Wise; when it clears, the files come by email as a zip, within one working day. A hosted checkout may replace the invoice later.
What's inside
- A 10-minute setup README
- TEAM.md: roles, how sessions message each other, and branch, port and process rules so two sessions don’t collide. It works without any messaging feature: the board is the source of truth and you relay messages by hand if you need to
- A BOARD template: one shared file where status, decisions and blockers live
- Handoff, review and QA templates: what to write when you pass work on, and what a reviewer must check
- GATES.md: green (just do it), amber (a human OKs it first), red (don’t)
- LESSONS.md in two parts: things that actually went wrong for us, told plainly, and rules we adopted early with no incident behind them
- scrub-check.sh: greps a folder for emails, key-like strings, home paths and .env files, and lists binary files it can’t check
Who it's for
- Solo founders and freelancers already paying for Claude Code who have started running two or more sessions
- People who want a starting structure to edit, not a framework to install
What went wrong (kept in on purpose)
LESSONS.md Part 1 lists what actually happened to us, for example: a relayed “the owner confirmed it works” that was false the same day; a reviewer-approved claim nobody had traced to code, found false after messages had gone out; a near miss where one pkill could have killed every session's build; two sessions hitting their usage limit by running four or five research agents at once. Each became a rule. Part 2 is rules we adopted early, labelled as having no incident behind them.
Made by Daksh Bathla, who also builds Itiner.
What it never does
- Not software or an app: it is markdown, plus one small shell script. Nothing to install or host
- Not “AI runs your company”. The sessions need a human who reads the board and says yes or no, and the files say so
- Not a guarantee of better code or lower cost. It gives you structure; the results depend on your project
- Not my real working files. Those hold client and product details, so this is a clean rewrite, with no real names, domains or data in it
- Not affiliated with or endorsed by Anthropic. It is built for Claude Code but is just text, and may not match future Claude Code features
Questions
Do I need Claude Code?
The files assume it. The ideas (a shared board, written handoffs, a separate reviewer) work with any tool, but the examples use Claude Code.
Does it work on Windows?
The files are markdown and work anywhere. The one script is for macOS/Linux shells.
Will you update it?
I’ll send fixes to everyone who bought it, by email. No promised schedule.
Is it only for big projects?
No. It starts paying off at two sessions and one repo.