# See Without Lauding _Date: 2026-03-19_ _Updated: 2026-03-19_ ___ So I've been messing with Claude Code (and other AI things) a lot more since January. After it goaded me to restart (or actually start) my blogging, I've decided to also note down my general feelings about this whole AI generation thing over here. Presumably updating as we go. Very off-the-cuff! ___ ## How I use Agents - I see AI as another tool to use. A very powerful one, and not just for coding. - SSH + tmux into a linux workstation for the most part, almost all Claude. - Each tmux session corresponds to a different Claude Code (CC) long-running session... mostly. - Mostly Claude Opus 4.6 main agents, Codex for some reviews. I want to try out Gemini and other AIs if only for blindspot reviewing. - Even a 10-agent code review team is not too much if you're feeling too lazy. I take care to try and prompt them slightly differently -- hoping that it would somewhat perturb their activation spaces out to prevent overlapping blind spots. - I attempt to anti-RLHF them by being direct, even quite rude! - I made up my own `[@macro]` system, basically lightweight "skills" you can `[@chain]` in a single prompt - Saves tokens (until microcompaction of tool use?) but more importantly: saves typing + can be explicit. - \[@commit] macro takes a file lock and auto-approves an all-in-one add+commit command wrapper, allowing multiple agents per worktree (which I don't use) - I have a `claude` wrapper changing `CLAUDE_CONFIG_DIR` for each sessionId/session name to prevent JSONL contamination cross-agents - Very important for long-running sessions in parallel -- But I don't know if the issue has been patched or not - Static analysis and similar tooling as much as possible -- still evolving! - Python: beartypes, ruff, basedpyright + hypothesis/mutmut - Sometimes agents will outright disable rules despite not needing to -- stop them, obviously. - Python tests are... really slow for some reason, compared to my scala munit tests. - Scala: ZIO, scalafmt, scalafix (almost all rules + no println), scalacOptions force more errors, scalacheck, pitest (...not pytest) ## Notes from my own AI usage? Take with less salt: - Claude likes to uncritically relay findings/ideas from its subagents. Tell it to stop being useless. - Claude seems to produce significantly better Scala FP code than Python. So much so that I've been using Scala for AI tooling. - I'd bet +1 standard deviation of code quality, probably. - Unless it tries to write Scala as if it were Python. - Wow, AI produces really questionable tests. Better to write tests first/in parallel with impl, and many people seem to have corroborated. Take with more salt: - Claude.ai degrades on complex problems much more than my Claude Code setup. - Opus 4.6 likes to ignore specific directions. Not sure about other models.