Built with Claude Code Study Portal

Key takeaways

that reads/writes files, runs commands, searches the codebase, and orchestrates sub-agents.

regression-tester, quality-gate, code-reviewer, chief-architect, visual-qa, tilt-ops), each with a focused role — and implementers are kept separate from the verifiers.

code-reviewer per domain → visual-qa + regression-tester (if UI/API changed) → an aggregated LIFECYCLE_PASSED / LIFECYCLE_FAILED report. Failures route back, get fixed, and only the failed check re-runs.

cleaner than you found it.

only real dependencies.

results are visible to the human and no duplicate dev servers spawn — agents never run raw npm/dotnet checks.

standards, and a live roadmap** — gives the agents continuity and a shared definition of "done."

output is always verified with real tooling. The leverage comes from tight verification loops, not blind code generation.

• The platform is built and maintained largely with Claude Code, an agentic tool
• Work is split across specialized subagents (frontend-dev, backend-dev,
• A mandatory lifecycle runs after every change: parallel quality-gate +
• Fix everything you find, including pre-existing issues — leave the codebase
• Parallelize independent streams and split the gatekeepers by domain; sequence
• The Tilt MCP feedback loop runs all verification through shared resources so
• Supporting machinery — **persistent memory, skills/slash commands, centralized
• Humans own direction, architecture, approval, and all irreversible actions; AI