AUTO-BUILDER FAQ

Plain answers, including the limitations.

What is AUTO-BUILDER?

An open-source Claude Code skill (/auto-builder) with nine subagents that runs a structured development lifecycle from a high-level goal: discovery, requirements, architecture, planning, implementation, verification, review and delivery. See the features.

Is it an official Anthropic product?

No. It is an independent community project and is not affiliated with or endorsed by Anthropic. The engineering practices it draws on (plan before coding, verify, delegate to subagents, keep project context in files) are inspirations from public guidance, not endorsements.

Which tools does it work with?

Claude Code only. It uses Claude Code skills and subagents. Other coding assistants are not supported.

Will it work unattended or in the background?

No. It runs inside your Claude Code session and stops at defined checkpoints. It does not run after the session ends, and because it is explicit-invoke only, scheduled or looped re-entry will not start it.

Can it push code, deploy, or spend money without asking?

Its instructions say it must ask first before deployments, spending, destructive data operations, exposing secrets, pushing, opening pull requests or changing product direction, and it never asks to bypass permissions. These are instructions to the model, not hard guarantees. Use Claude Code permissions and sandboxing as the enforcement layer and review the diff.

Does it change my git history?

It works on a new auto-builder/<slug> branch and makes local checkpoint commits by explicit path. It is instructed not to rewrite history, force-push or commit your uncommitted changes. Tell it you don't want a work branch to skip that step.

How much does it cost?

AUTO-BUILDER itself is free under the Apache-2.0 license. A run uses your normal Claude Code usage and subagents add to it. There is no cost estimate here because none has been measured.

How is it different from just prompting Claude Code?

It packages a repeatable process: a plan and acceptance criteria before code, a persistent task file that survives compaction, milestone gates with recorded evidence, bounded fix loops with root-cause analysis, independent review at meaningful milestones, least-privilege subagents and git rules. Whether that yields better results for your project is for you to judge; no benchmarks are claimed.

Does it need Python?

No. The skill and agents are Markdown. An optional helper script (scripts/ab.py, standard library only) can check task tables, compare recorded state with git, and format-check browser QA evidence. If Python 3 is missing the lead applies the same rules by hand. The helper writes nothing, uses no network, is not wired into any hook, and may trigger an ordinary permission prompt when run.

Does it test my app in a real browser?

Only if a browser is available: the project's own end-to-end tests, browser tools in your Claude Code session, a runner such as Playwright installed in your project, or headless Chrome. AUTO-BUILDER ships no browser. If none is available the report says "browser QA not performed" and the affected journeys are untested. The evidence file has a checker that rejects inconsistent reports, but it proves nothing about what the app did.

What are the known limitations?
Does it support Windows?

A PowerShell installer is provided but untested by the author. The optional hook scripts and the template helper are bash, so use Git Bash or WSL for them.

What licence is it under, and can I contribute?

Apache-2.0. Contributions are welcome; see CONTRIBUTING.md and open an issue for larger changes.

Ready to try it? Install AUTO-BUILDER.