Use AUTO-BUILDER: a goal-driven coding agent in practice

Invoke it with /auto-builder followed by a goal and any constraints. It is explicit-invoke only.

Invoke the skill

/auto-builder Add a rate-limited, paginated /v1/orders API with tests. Reuse the existing stack. Do not deploy anything.

With no argument it inspects the repository and asks for one concise goal. Good goals state the outcome, constraints and what not to do. Useful phrases: "stop at the plan checkpoint", "no branch".

How requests are routed

The lead first checks whether the request is trivial (a single file, no design, data, security or user-visible decision); if so it just does it and verifies it. Otherwise it restates your goal, classifies it (for example bug fix, new feature, refactoring, performance, security assessment, code review, research, documentation or release), tells you which playbook it chose in one sentence, and runs it. Say route as <category> to override. A new project runs the full lifecycle; every other category runs a shorter playbook inside one milestone of the task file, under the same limits and gates.

Example goals

CategoryExample
AI/ML appAdd semantic search over support articles with an evaluation set measuring recall@5
Multi-agent systemBuild a triage bot that classifies GitHub issues and drafts replies for human approval
Full-stackBuild a team habit tracker with auth, a REST API and a React UI
FrontendMake the settings page keyboard accessible and responsive
BackendAdd rate-limited, paginated endpoints with OpenAPI docs and tests
Developer toolCreate a CLI that lints OpenAPI specs against our style guide
AutomationAutomate weekly invoice export from a drop folder, with a dry-run mode
Data engineeringBuild an idempotent daily ETL into Postgres with quality checks

Tutorial: your first run

  1. Install the skill (guide) and open Claude Code in a project, preferably a git repository with a clean working tree.
  2. Run /auto-builder with a small, concrete goal for your first try.
  3. Reconnaissance. The lead reads your repository, notes git state and finds the test, lint and build commands. If you have uncommitted changes it records which files they touch so it never commits them.
  4. Discovery to planning. It frames the problem, writes requirements with testable acceptance criteria and a task list under docs/auto-builder/ (you can run bash .claude/skills/auto-builder/scripts/init-artifacts.sh yourself to see the templates). TASKS.md has a Status block, an acceptance checklist with an evidence column and an iteration budget.
  5. Checkpoint A. If a consequential choice is involved, or you asked for it, it pauses and summarises direction, architecture, risks and assumptions. Approve or redirect.
  6. Build, verify, review, one milestone at a time. It implements vertical slices on an auto-builder/<slug> branch, then runs the milestone gate: tests, regression checks, real flows, and a separate review by critical-reviewer (and qa-engineer for user-facing behaviour) at meaningful milestones and at the end. For projects with security risk (authentication, user data, payments, APIs, uploads, secrets, AI tools) security-engineer also threat-models the design and reviews relevant milestones; a High finding is the one case where it will ask you to accept residual risk. It records evidence in TASKS.md and its log in PROGRESS.md. Failures are root-caused, fixed, and re-checked, within the bounded loop.
  7. Delivery report. You get what was built, decisions, checks that actually ran, which criteria are verified, partially verified or untested, whether the review was independent, limitations, run and deploy guidance, and files changed. Read it, then review the diff yourself.
  8. Merge when satisfied. It is instructed not to push or open a pull request unless you explicitly ask.

Tips

Questions? See the FAQ or the architecture.