I hated making this video...

TL;DR

• 1

  Claude Code skills can execute scripts at load time: Theo argues this makes skills like his "repo explorer" far more reliable because the model can inspect cached repos immediately instead of wasting turns discovering state.

• 2

  Claude MD imports are a clever workaround for the Claude MD vs AGENTS.md split: Using @path/to/import, he can make CLAUDE.md include AGENTS.md plus Claude-specific instructions, without symlinks or duplicate files.

• 3

  Workflows are the most impressive and most expensive feature: In his test, Claude Code wrote 240 lines of throwaway JavaScript to orchestrate a 15-agent PR audit, and usage jumped from 9 percent to 24 percent in minutes.

• 4

  The real magic is 'code mode,' not just more tool calls: Theo's point is that when the model writes code to define sub-agents, prompts, schemas, and stages, it creates custom orchestration logic on the fly that fixed workflow UIs cannot match.

• 5

  Small UX touches matter more than people admit: Theo calls out /side for asking parallel questions mid-task, local-only claude.local.md, worktree support, rewind/branch, fullscreen terminal mode, and remote session viewing from the Claude app.

• 6

  He is praising Claude Code so competitors will steal its best ideas: Theo says plainly that this is not a glaze piece for Anthropic, but a wishlist for Codex, Aider, OpenCode, and others to copy what Claude Code uniquely gets right.