I kept doing the same thing manually: write a spec, have Claude implement it, review the code, fix issues, reflect, commit, repeat — resetting the context window by hand at each step. So I automated the loop.

cc-pipeline takes a BRIEF.md describing what you want built and orchestrates Claude Code through the whole SDLC: spec → research → plan → build → review → fix → reflect → commit. Phase by phase, overnight, while you sleep. Each step gets a fresh context window via the Claude Agent SDK (I tried tmux + send-keys first — do not recommend).

I've run it on a dozen projects: an Elixir port of an existing open-source gateway, a Trello-style kanban board twice (Astro and Rails), a statistical analysis in R (I've never written R), Tetris twice from the same brief, and a Kirby platformer attempt (that failed).

Full write-up of the experiments here: https://curiousagents.substack.com/p/experiments-in-building...

Quick start:

``` cd your-project npx cc-pipeline@latest init ```

Then open Claude Code and ask it to help you write the BRIEF.md. Then `npx cc-pipeline run` and walk away.

The steps are defined in .pipeline/workflow.yaml — you can add, remove, or reorder them. The prompts are plain markdown. Easy to tune to your own SDLC.

Would love feedback on the approach, especially from anyone who's tried similar things.