Skip to content

Blueprint

Which version? The large 0.8 changes (ships in 0.8) are still moving. See Which version should I use? before you start.

The big picture: Blueprint is a delivery lifecycle for products that Claude Code builds on GitLab. You clone it as your project’s first commit. It takes the project from configuration through spec, plan, build and release, with review gates sized to each change and a ledger that records what each gate actually caught. It contains no application code and picks no language or framework.

Why it matters: AI coding agents are fast, but speed without structure makes messes faster. Most projects only add review and release discipline months in, after the incidents that prove they need it. Blueprint gives you that discipline on day one.

The bottom line: You describe what you want. The workflow supplies project context, runs the reviews that change actually needs, and lands the result as a reviewed merge request with tests and a changelog entry already in place.

New here? The lifecycle shows the whole loop on one page: what to type at each step, what done looks like, and how to get back in.

Every page opens with the essentials: The big picture (what it is), Why it matters (what it buys you), then short bullets. Skim those and move on, or keep reading — the Go deeper sections below them hold the full detail and the reasoning.

New to agents, skills, gates, hooks, or worktrees? Keep the Glossary open in another tab.

  • Without: you ask for a change. The agent writes code, maybe writes tests, maybe touches something you didn’t ask about, and you reconstruct what happened from the diff.
  • With: the change starts from a tracked issue, gets its own branch, goes through the reviews sized to it, and arrives as an MR that records which checks ran and what they found.

Blueprint treats Claude Code as a contributor working inside a defined process — the same branch, review, and test discipline a human contributor follows, written down once instead of re-explained every session.

  • A merge-request-only workflow. Branch, commit, and MR conventions are written down and enforced by Git hooks and CI.
  • Specialist agents. Reviewers for security, permissions, performance, regressions, database changes, UX, accessibility, and more — each run only when a change needs it.
  • Skills. One-command workflows for setup (/kickoff), planning (/dotplanning), shipping (/mr), and releasing (/release).
  • A CI pipeline with governance checks that work out of the box, plus add-ins for Python, Node.js, Go, Docker, and Helm.
  • Gates that prove they still work. Every check script can demonstrate it still fails on a bad input, so a green result means something.
  • Safe parallel work. scripts/wt gives each issue its own folder and branch, so several Claude sessions can run at once without colliding.
  • A fragment-based changelog and a release script that assembles it.

You supply only what nobody else knows — the project’s name, its users, its stack, what version one must do, and setup choices such as the license and the CI add-ins. /kickoff asks for those in a few batched rounds, writes them in, and lists what it assumed.

You are… Fit
Curious about AI coding but haven’t built anything yet Probably too early — see the maturity ladder
Have built a small app with Claude Code Good next step
Moderately technical and comfortable with Git Good fit
Building a serious personal project you’ll maintain Very good fit
A professional developer using AI agents Strong fit
A team using Claude Code and GitLab together Strong fit
Looking for a no-code tool Not the right tool
Wanting a quick one-off prototype Probably overkill

Between the lines: Blueprint is built for the move from “Claude helped me make this” to “I have a repeatable way to build and maintain this.”

It’s a later rung, not a starting requirement. Adoption runs through four maturity levels: Prototype, Controlled, Hardened, and Measured. You move up because something hurt, not because the next level sounds better. Each level has a “you are here when” test, the mechanism that fits it, and a “leave when” symptom; see the maturity ladder. Blueprint itself is a Level 2 template.

Required: basic command line; basic Git (clone, branch, commit, pull, merge); basic GitLab; access to Claude Code; a project you want to apply this to.

Helpful, not required: some programming; what a test suite does; basic CI; reading Markdown and YAML; experience directing an AI coding agent.

Not needed: professional development experience, building AI agents, CI/CD expertise, GitLab administration, or prompt engineering. Blueprint teaches what applies as you go.

  • Not a model or a Claude Code replacement. It configures Claude Code; it doesn’t replace it.
  • Not an app generator or scaffold. No src/, no framework, no Dockerfile. Add those the way you normally would.
  • Not a guarantee. It makes AI-written code reviewed, not automatically correct, and it doesn’t replace understanding what you’re building.
  • Not a dependency. Once cloned, the files are yours. Nothing pulls updates from Blueprint.
  • Not GitHub-first. CI, issue templates, and MR templates target GitLab. The scripts also work with gh; see Removing what you don’t need.