AI Workflow
For the past five months I have been using terminal-based AI tools not to generate code faster, but to make my development process more systematic and verifiable. These are the three disciplines I built around AI-assisted development.
I do not consider a task complete just because code was produced. I run the service, check the output, add tests when possible, and confirm the change did not break the existing system. Recurring mistakes get converted into project rules, checklists, or automations so they do not happen twice.
Built a rule pipeline: errors are logged once, and a second occurrence turns them into a reusable skill with trigger conditions, fix steps, and verification.
I work across Hermes, Claude Code, and Codex daily. I learned the hard way that spreading a task across tools causes context loss and repeated work. My workflow defaults to a single agent with shared project rules and documentation, and only parallelizes on genuinely independent tasks.
Maintains a cross-agent knowledge base (AgentMemory) and shared project conventions (AGENTS.md) so every tool starts from the same context.
To make AI tooling sustainable, I built TAP, an open-source CLI that tracks model usage, token consumption, and unnecessary spend across sessions. I route daily tasks to economical models and reserve stronger models for planning and complex debugging.
TAP analyzed 800+ sessions and surfaced $2.5K in identifiable overpay across $9K of tracked spend.
Why it matters
I position AI as a practical assistant that automates flows and creates smarter experiences, not as hype. The discipline is simple: understand the problem, break it down, use AI where it accelerates, and verify everything it produces before it ships.
This is the difference between vibe coding and engineering with AI. Both produce code. One produces code you can explain, maintain, and ship with confidence.