What I use AI for
I use AI for software, automation, and technical decisions. I want small changes I can inspect, evidence I can check, and handoffs I can pick up later.
What I build
I have not built a standalone product I consider significant enough to show yet. I am trying to make this setup cover the work of at least two senior developers on my team: data integration, orchestration, data engineering, and data quality. I am also testing how my experience transfers to a larger company where many engineers are not AI-native yet.
Agents and runtimes
- Most of the actual work happens in Codex. It handles planning (astra), tool use, implementation (luna:high-max), debugging(sol:medium), and review(sol:medium).
- I use subagents for independent, bounded pieces of work. The main thread keeps the decisions, joins the pieces, and does the final verification.
- I use Claude Code occasionally, mostly for chat, thinking, and conceptual work.
- Hermes is mostly for personal work and experiments: agent-to-agent work, trading and backtesting, Telegram bots, and conversational workflows.
Skills and guardrails
I generally dislike prepackaged skills. Superpowers and Ponytail are exceptions: they help me structure the work and cut slop. I built most of the rest myself because generic prompt packs rarely fit my needs. About 95% of my Hermes skills are tagged “learned.”
My process is straightforward:
- start with the relevant instructions and the smallest useful context;
- reuse existing helpers and keep changes narrow;
- use isolated workspaces for meaningful source work;
- debug from the root cause, not the visible symptom;
- test the smallest meaningful behavior first, then broaden checks before handoff;
- verify the result before calling it done.
For live behavior, I keep source code, tool output, and what actually happened separate. Tests tell me about code paths. I still check the result in the environment where it matters.
How the tools fit together
The terminal and its tools extend Codex when I need to inspect or change things. I use the other tools for their own purposes rather than treating them as one system.
Machines and terminal
For work on Windows, I use WSL, a terminal, and zsh. For personal work on macOS, I use Ghostty and zsh. My main machines are a Lenovo P14 and a MacBook Pro with an M3 chip.
How I work on a task
1. Set the exact target, boundary, and success evidence.
2. Read the relevant instructions, source, tests, and context.
3. Preserve unrelated work and fix the shared root cause with the smallest change that works.
4. Run focused checks, then broaden to related tests, lint, and diff checks. For operational work, check the actual tool or service involved.
5. State what is known and unknown, then leave a short durable note when it will matter later.