Automating Myself Out of Development
6 days ago
- #AI-assisted development
- #workflow automation
- #Claude Code
- The author describes a journey of progressively automating their development workflow using Claude Code, starting with local synchronous sessions and evolving to a fully automated, scheduled system.
- Initial phases involved handling multiple terminal tabs and windows for parallel feature development, leading to context-switch fatigue and a need for better delegation.
- The workflow moved to an isolated EC2 instance to reduce security risks and project cross-contamination, though it initially felt slower due to missing conversational context.
- The author implemented a daemon-driven system using GitHub issues as a planning board, with labels representing states and automated steps for enrichment, brainstorming, spec creation, implementation, and review.
- Human touchpoints were strategically placed at key stages: reviewing enriched issue bodies, approving or editing specs, reviewing plans, and merging, while failures trigger a 'needs-attention' label for manual intervention.
- The current system includes automated context-gathering and brainstorming with logs for transparency, aiming to delegate implementation without delegating critical thinking.
- Challenges remain, including uneven test quality, architectural drift, and the need for better categorization of features and more sophisticated review agents.
- The author cautions against fully delegating thinking to AI, noting that it increases technical debt and emphasizes the continued importance of code review, static analysis, and security audits.