Open-source software factory for Claude Code
Three skills turn rough ideas into reviewed pull requests—while keeping product decisions and every merge in human hands.
The operating system
Automation handles the repeatable middle. People define what good looks like, release work into the queue, and decide what ships.
Start with a rough thought, not a polished ticket.
HumanThe agent researches the repo and removes ambiguity.
/finn-specYou apply agent-ready when the contract is right.
Human gateOne approved issue becomes one tested pull request.
/finn-buildThe PR is checked against the issue and green CI.
/finn-reviewA person reads the evidence and makes the final call.
HumanThe three skills
Each skill owns one stage and has a hard stopping point. That is what makes the system legible—and safe to leave running.
/finn-specIt reads the existing code before asking questions, then interviews you until two engineers would build the same observable behavior.
/finn-buildIt claims the next safe issue, protects unrelated work, implements only the acceptance criteria, runs the project checks, and opens a PR.
/finn-reviewIt reviews the exact commit against the linked Linear issue, required GitHub checks, mergeability, defects, security, and scope drift.
Start here
Finn-loop installs into the repository where you want the factory. You need Claude Code, Linear, GitHub, and project CI.
Open Claude Code in your target repository and paste the setup request. It checks your team, labels, permissions, default branch, and required checks.
Set up Finn-loop from https://github.com/finna/Finn-loopStart with something useful but small. The spec agent reads the codebase first, then asks only the product decisions it cannot answer itself.
/finn-spec Add team invitations to settingsRead the drafted Linear issue. When the acceptance criteria, non-goals, and verification steps are right, apply the one approval label.
agent-readyapplied by youRun the builder and reviewer in separate Claude Code sessions. They self-pace, stop when their queues are empty, and leave every merge to you.
/loop /finn-build/loop /finn-reviewWhy it stays sane
The agents are capable. The factory works because it narrows their freedom at the moments where confidence can become expensive.
If it is not in Linear, it does not exist. The queue—not chat history—is the source of truth.
A blocked issue leaves the automated queue until a person answers the exact question and removes the label.
A review comment cannot expand the issue. Acceptance criteria and non-goals change only in Linear.
No required checks means no automated approval. Red or pending checks never count as green.
The assignee is a cooperative lock. Run one builder loop per Linear team to avoid same-user races.
Every irreversible step stays with a person. Agents never merge and never enable auto-merge.
A normal day
Spend fifteen focused minutes on the places where your taste and context matter. Let the loops handle the repeatable work between those decisions.
Release only the exact contracts you want built.
Make the product decisions the agents refused to guess.
Read the verdict, checks, risk, and manual test steps.
Good questions
No. The builder never merges or enables auto-merge. A loop-approved label gives you structured evidence for your own merge decision.
The builder posts one concrete question, applies the blocked label, unassigns itself, and moves on. The issue returns only after a human answers and removes blocked.
Use one builder loop per Linear team. Two sessions authenticated as the same Linear user cannot reliably lock each other. Separate teams and clean worktrees can operate independently.
It runs the checks your project already defines and requires GitHub evidence before automated approval. Your repository still owns its test suite, build, security, and deployment checks.
Ready when you are