I was an "old school" programmer back in 2025. I remember Cursor agents launched around July 2025, and they felt like mere entertainment, as they made a lot of errors and basic mistakes that a normal dev would never make. I never imagined at the time that a for loop over an LLM API call would mature into taking complete ownership of large parts of codebases, to the point where people now trust it enough to let it run autonomously and drive impact.
When I first got access to GitHub Copilot back in June 2025, I absolutely loved developing with it. It could generate tons of boilerplate and documentation, which meant I could spend my time purely thinking. The ability to jump straight into a problem was a privilege. Now, I don't even start thinking without letting Claude cover the basics for me, research the problem, and propose things.
I started using Claude in January 2026 and Claude Code a few months later. A 100-foot tsunami of AI psychosis hit me when I first started agentic engineering. I used Claude Code like a "Solve All My Problems Shotgun": I used to one-shot implementations and push them. I did not care about what was in them, and my only source of verification was that the frontend showed the correct data and did not throw 500 errors! After struggling for months, here is how I improved my usage of Claude Code, or any other harness for that matter, and gained back some degree of job satisfaction.
My Environment
Before explaining how I use Claude Code, I'll explain where I use it.
I use it inside my terminal. I find Ghostty to be the best terminal emulator: it is blazing fast, responsive, never crashes on me, and unlike the standard GNOME terminal, it doesn't fuck up outputs.
Inside Ghostty, I use Herdr. It is simply the best for viewing and managing multiple agents and terminal windows. It is unique because it gives me just enough context about the running status of my agents. It does not show PRs, worktrees, file diff counts, or annoying popups like most GUIs do, which helps me stay focused.
Finally, and obviously, I use Claude Code in the CLI. It is a bit slow, and its startup time is longer than I would like. Compared to Codex, it is extremely slow, but it is still much faster than the desktop app, at least in my experience. Whenever I try out GUIs, I find myself drifting back to my terminal, always.
My Global AGENTS.md
Here is my global AGENTS.md.
These instructions make Claude output information exactly how I want it. I hate it when it just spews out a wall of text where 60% of it is complete noise.
My instructions:
- First principles
Think from first principles. Reason from underlying constraints, mechanisms, and invariants rather than relying on convention by default.
Pretty self-explanatory. I want my agents to think from first principles.
- Concise, STE100 language
Be concise and to the point. Strictly use ASD-STE100 language specification when generating text/explanations.
I picked up this instruction from Andrej Karpathy, and it works insanely well. It does not beat around the bush, and it does not give me a senseless wall of text anymore.
- Visual explanations
Prefer visual explanations: For any explanation involving multiple steps, components, branches, or interactions, include a relevant diagram. Use ASCII data-flow diagrams to show labeled data, components, transformations, and flow direction; ASCII sequence diagrams to illustrate communication between entities, including participants, message direction, and order; and call-stack diffs when proposing function or class changes, showing before-and-after caller → function/method → callee paths with added, removed, or rerouted calls clearly marked. Use text alone only for simple facts or single-step actions.
I am a visual person. I literally cannot understand text, and I especially cannot understand architecture through text alone. I absolutely need visual explainers.
My Project-Local AGENTS.md
Here is my project-local AGENTS.md.
It is also pretty self-explanatory.
One thing I am absolutely allergic to is bloated AGENTS.md or CLAUDE.md files. I physically cringe looking at these files when they are over 200 LOC and the devs think they are doing right by the agent by dumping every single thing about the codebase into context on the first turn. I hate it.
Skill Files
Here are the skill files I use:
- Ponytail: for concise and maintainable code
- Remotion: for generating videos for the project
- Cloudflare and AWS: to work with CF & AWS via agents
And that is it! No pstack, gstack, or Matt Pocock skills. I like to keep it simple, and whenever I really want a particular functionality, I temporarily provide that skill and then purge it after its use is complete.
Conclusion
So that was it! This is my workflow. It may not apply to everyone, but this is how I reduced fatigue while working with AI and actually started having fun again.
Top comments (0)