Project Initialization and Tech Stack #
- Git for Version Control: Employs Git as the foundational layer for all projects to ensure proper version control and seamless integration with AI agents.
- Better T Stack: Uses this tool to scaffold projects, allowing for a customized, performant stack (e.g., Next.js, Polar for payments, Convex for backend, and Bun for runtime).
- ShadCN Create: Utilizes this to generate design templates, streamline styles, and maintain visual consistency across the application.
Pair Programming & Local Development #
- Kilo Code: Serving as the primary VS Code pair programmer, it is used for small refactors, explaining codebases, and quick manual changes.
- GLM Coding Plan: Prefers using this within Kilo Code due to its cost-efficiency and high performance.
- API Flexibility: Recharges Kilo accounts to swap between different model APIs depending on the specific coding need.
Background Coding & Task Delegation #
- Verdent: Used for complex task delegation. It operates by creating isolated workspaces or branches to protect the main repository from errors.
- Features a $200/month plan for high-volume tasks.
- Noted for high memory efficiency and a robust notification system for task completion.
- Conductor: Employed for small to medium changes to save Verdent credits.
- Optimized for 24/7 operation at high speeds.
- Recommended as a superior, lower-cost alternative to Claude Sonnet for general agentic tasks.
Pull Request (PR) Management and Code Review #
- Graphite: Used to manage the Git workflow and pull requests. It supports "stacked merges," allowing multiple PRs to be merged simultaneously while resolving conflicts in one go.
- Gemini Code Assist Review: A free, automated tool that provides line-by-line annotations and reviews for every PR created.
- Grapile: An alternative review tool noted for speed and quality, though currently requiring manual mentions on GitHub to trigger.
Preferred AI Models #
- GLM 4.7 & Opus 4.5: Identified as the top-performing models for the current coding workflow.
- Human-in-the-Loop: Prefers active monitoring of agent progress over fully autonomous "loops" to ensure code quality and manage costs effectively.
Summary #
The 2026 AI coding workflow focuses on a highly modular approach where the developer acts as an orchestrator rather than a manual coder. By utilizing Better T Stack and ShadCN for setup, Kilo Code for immediate pair programming, and agents like Verdent and Conductor for background task execution, the developer can ship features rapidly. The workflow is finalized through automated rigors using Graphite for PR management and Gemini/Grapile for autonomous code reviews, ensuring that the developer only intervenes for final quality checks and merging.
last updated: