Paste this into Claude Code — hit the copy button, paste, and it scaffolds your hub:

Create a project hub from this template: https://github.com/PakhomovAlexander/project-hub

Project Hub — a template for running a project with Claude Code

A Project Hub is one repo that becomes the cockpit for your project: shared language, the plan, the decisions, the live status, and the controls — all in one place an agent and a human can both operate from. The product code stays in its own repo(s); the hub links them in and drives planning, review, and ops from above.

This template is a battle-tested shape for that cockpit, designed so you can hand it to Claude Code and have it scaffolded for you in minutes.

The 30-second pitch

Most "AI in the repo" setups give the agent a CLAUDE.md and hope. A hub goes further:

• CONTEXT.md — a real glossary (ubiquitous language) so the agent uses your words.
• CLAUDE.md — the working agreement: invariants the agent must never break, plus PR/CI, verification, and doc-honesty discipline.
• AGENTS.md — a vendor-neutral pointer to CLAUDE.md, so agents beyond Claude Code (Cursor, Codex, …) drive the hub from the same rules — one source of truth, no drift.
• docs/adr/ — decisions recorded with options + consequences, superseding over time.
• docs/tracker.md — a living status board: what's true right now, dated.
• docs/workstreams/ — deep design docs for in-flight work, with "resume here" sections.
• docs/service-catalog.md — a tiered, agent-facing map of every service/repo + doc status, so a session looks a thing up instead of re-deriving it.
• scripts/repos.sh + repos.manifest — link many live repos into repos/ as symlinks.
• scripts/worktree.sh — run several agents over one hub at once, each in its own git worktree, without branch/index collisions.
• .claude/ — a safety hook that prompts before prod/destructive commands.
• .github/ docs CI — markdownlint + an offline link check on every PR, so the docs can't silently rot.

The result: an agent can pick up cold work, speak your domain, respect your rules, not nuke prod, and run alongside a dozen of its peers — and a human can read the whole project's state from one folder.

Use it (the "send a link" flow)

Point Claude Code at this repo and say what you want:

"Set up a project hub for Acme using this template: <link-to-this-repo>. It's a real-time collaboration app across three repos in the acme-inc org."

Claude reads SETUP.md — the agent runbook baked into this template — then interviews you for the specifics (name, repos, invariants, environments, team, which commands should prompt before running), generates a customized hub in a new directory, wires up the repo links, and verifies it.

You can also drive it locally:

git clone <link-to-this-repo> project-hub-template
cd project-hub-template
# then, in Claude Code:  "Follow SETUP.md and scaffold a hub for <my project>."

What you get

your-project-hub/
├── CONTEXT.md              # shared language / glossary — read first
├── CLAUDE.md              # working agreement + invariants for agents
├── AGENTS.md             # vendor-neutral entry point → CLAUDE.md (any agent)
├── README.md             # the hub's own front door
├── TEAM.md               # people ↔ GitHub ↔ ownership (optional)
├── Makefile · scripts/   # link/clone/status the repos · worktrees for parallel agents
├── repos.manifest        # the list of repos this hub coordinates
├── .claude/              # settings + the prompt-before-risky-commands hook
├── .github/workflows/    # docs CI: markdownlint + offline link check
├── .markdownlint-cli2.jsonc # light, high-signal Markdown rules
├── docs/
│   ├── index.md          # map of all docs
│   ├── plan.md           # the master plan (scope, workstreams, risks, decisions)
│   ├── tracker.md        # live status board
│   ├── service-catalog.md# every service/repo + doc status + the doc standard
│   ├── issue-lifecycle.md# how the backlog moves
│   ├── parallel-agents.md# running several agents at once via git worktrees
│   ├── adr/              # architecture decision records (+ a seed ADR + a template)
│   ├── workstreams/      # in-flight work: design + acceptance + resume-here
│   └── repos/            # one Tier-1 reference per linked repo
└── repos/                # symlinks into your real clones (gitignored)

Single-repo project? The setup drops the linking machinery — you keep the docs, the invariants, the ADRs, and the safety hook.

What's in this template repo

• SETUP.md — the runbook an agent follows to scaffold a hub (both AGENTS.md and CLAUDE.md point here). Read it to see exactly what the agent will do.
• template/ — the hub skeleton (placeholders + inline guidance) that gets copied and customized into your new hub.
• scripts/verify-hub.sh — run it against a generated hub to catch leftover placeholders, broken internal links, and a non-executable safety hook before calling setup done.

Why it works

The patterns here came from running a real multi-repo product launch entirely through a hub. The agent could resume cold work, respect hard invariants, keep a truthful status board, and operate across a dozen repos without losing the plot — because the cockpit made the project's language, decisions, and state first-class, not tribal knowledge.

This template is generic and project-agnostic. It contains no details from any specific project — just the shape. Make it yours.

License

MIT. Use it, fork it, share it.