Features: an autonomous Claude Code workflow with guardrails

Everything below exists in the repository today. Nothing here is a benchmark or a performance claim.

On this page: Lifecycle · Task router · Quality protocol · Security gate · Subagents · Safety controls · Project categories · Installers · Hooks

Nine-step autonomous development lifecycle

  1. Phase 0: Reconnaissance. Inspect the repository, git state, conventions and the commands that build, test and lint it.
  2. Phase 1: Discovery. Frame the real problem and users, research where facts matter, compare approaches, recommend one.
  3. Phase 2: Product definition. Write testable requirements, user journeys and acceptance criteria.
  4. Phase 3: Architecture. Choose the stack, design components, data and APIs, and record trade-offs; threat-model security-sensitive designs.
  5. Phase 4: Execution planning. Define measurable acceptance criteria and milestones with verification gates in a persistent task file, with a per-milestone iteration budget; add security controls where risk warrants.
  6. Phase 5: Implementation. Build one milestone at a time as vertical slices on a work branch; the lead commits only after the milestone passes its gate.
  7. Phase 6: Independent verification. Run the milestone gate: tests, linters and regression checks, real user flows, then an independent review at meaningful checkpoints.
  8. Phase 7: Iterate. Repeat plan, build, test, review, fix per milestone, with root-cause analysis of failures, until the completion standard is met or a blocker needs you.
  9. Phase 8: Delivery. Report what was built and, per criterion, what is verified, partially verified or untested, whether review was independent, the security status, and what you must do next.

The loop in phase 7 is bounded: three attempts per failing check and an iteration budget of five passes per milestone, recorded in the task file, so it is instructed not to run forever. Only you can accept a criterion as out of scope.

Quality protocol: milestone gates and evidence

Quality, not speed, is the stated objective. The skill instructs the lead agent to:

This is an instruction system for the model, not a guarantee that the result is defect-free. Review the work yourself.

Task router and playbooks

You do not have to push every request through the full project lifecycle. After an explicit /auto-builder, the lead classifies the request into one of twelve categories (new project, new feature, bug fix, investigation, refactoring, performance, security assessment, code review, research, testing, documentation, release) and runs a short playbook for it. For example a bug fix is reproduce, diagnose the root cause, write a failing test, fix, run regression, review. Trivial changes (one file, no design, data, security or user-facing decision) skip the ceremony. Routing can only add gates, never remove one, and text pasted from issues, pull requests or web pages cannot change a gate. When classification is ambiguous and consequential the lead asks one focused question, and never asks in non-interactive runs.

Around the router: product interrogation for substantial work (problem, user, status quo, smallest valuable solution, assumptions, success metrics, what not to build) with explicit skip rules; an architecture arena that is cheap by default and independent only for hard-to-reverse or security-critical decisions; dependency-aware parallel execution with owned files and a post-wave check that workers stayed inside them; fresh-context review (a separate context, honestly not independent of the lead's framing, and never called multi-model unless two models ran); bring-your-own-browser QA that says "not performed" when no browser exists; session recovery that verifies recorded state against git instead of trusting it; and opt-in, approval-gated pull request and release help.

What is checked and what is not. State files, issue and pull request text, web pages and tool output are treated as untrusted data that can never grant an approval or change a limit. An optional, read-only Python helper validates task tables and ownership, compares recorded state with git, and format-checks browser evidence. It writes nothing and proves no behaviour. Everything else here is an instruction to the model that the repository's tests check for presence and consistency, not for effect; none of it has been run live yet.

Security gate

For projects that handle authentication, user or personal data, payments, externally reachable APIs, file uploads, secrets, integrations or AI/LLM tools with side effects, the skill instructs the lead to involve security-engineer, a read-only, independent reviewer:

Findings are rated Critical, High, Medium or Low, with file and line evidence, an exploitation scenario and a minimal fix. A Critical finding blocks completion and cannot be accepted. A High finding must be fixed or explicitly accepted by you; an accepted High stays recorded as open residual risk, never as fixed. Medium and Low findings are tracked. A finding counts as fixed only with evidence from a re-check. Reports state whether they rest on static inspection, automated scanning or penetration testing (not performed unless explicitly authorized), and a clean scanner run is never reported as "secure". The agent only reads and reports: it does not edit files, touch production, run destructive tests or access external targets, and these limits are instructions, not technical guarantees.

Nine Claude Code subagents

Each subagent is a Markdown file in .claude/agents/ with valid Claude Code frontmatter. Subagents start with no conversation history and cannot ask you questions, so the lead gives each a precise brief and integrates its output.

The nine AUTO-BUILDER subagents (generated from .claude/agents/)
AgentResponsibilityToolsCan edit files
product-managerClarifies product goals, target users, user journeys, scope, prioritization and measurable acceptance criteria. Use for discovery and requirements, not coding.Read, Grep, Glob, WebSearch, WebFetchNo edit tools
researcherResearches current libraries, APIs, standards and competitor patterns from primary sources. Use when current external facts materially affect a decision.Read, Grep, Glob, WebSearch, WebFetchNo edit tools
architectDesigns or independently critiques system architecture, data models, interfaces, security boundaries and implementation plans. Read-only.Read, Grep, GlobNo edit tools
frontend-engineerImplements and reviews frontend features: components, state, responsive layouts, accessibility and browser interactions.Read, Edit, Write, Bash, Grep, GlobYes (assigned files)
backend-engineerImplements backend services, APIs, database schemas, migrations, integrations, validation and data pipelines.Read, Edit, Write, Bash, Grep, GlobYes (assigned files)
ai-engineerDesigns and implements AI/ML features: LLM integrations, RAG, agents, evaluation, prompts and data/model pipelines.Read, Edit, Write, Bash, Grep, Glob, WebSearch, WebFetchYes (assigned files)
qa-engineerDerives tests from acceptance criteria, runs verification, explores edge cases and independently validates behaviour. Read-only auditor: reports findings and proposes tests.Read, Bash, Grep, GlobNo edit tools
security-engineerIndependent application-security reviewer: threat modeling, attack surface, OWASP Top 10, authn/authz, injection, XSS/CSRF/SSRF, secrets, dependencies, data privacy, AI/LLM security. Read-only; reports evidence-based findings.Read, Bash, Grep, GlobNo edit tools
critical-reviewerIndependently reviews plans, diffs and implementations for correctness, security, regressions, maintainability and missed requirements. Read-only.Read, Grep, Glob, BashNo edit tools

The reviewer, QA and security agents also have Bash to run checks. Claude Code cannot restrict Bash by argument, so they are instructed, not technically prevented, from modifying files.

Safety and approval controls

Project categories

AI/ML apps · AI agents and multi-agent systems · full-stack · frontend · backend/API · developer tools · research · automation · data engineering. A category guide tunes discovery, architecture and verification per type.

Safe installers

install.sh (Linux, macOS, WSL, Git Bash) and install.ps1 (Windows PowerShell) install for one project or all projects. They support --dry-run, never overwrite your files unless you pass --update (which keeps numbered backups), refuse to replace symlinks, and report stale files from older versions. See the installation guide.

Optional hooks

Two opt-in hook scripts are included but never enabled automatically: protect-secrets.sh blocks edits to secret-bearing file paths, and verify-on-stop.sh runs your check command before Claude finishes, with bounded retries and a timeout. Limits: the first does not cover Bash or Read, and the second is a strong nudge rather than an absolute gate.