▶ YouTube Summaries

• https://www.youtube.com/watch?v=FDxW2bfBOWE
• Original title: I hated making this video...

A reluctant, hat-on appreciation of the features Claude Code does better than every other agentic-coding harness. Theo's explicit goal isn't to sell Claude Code but to get codex, opencode, pi, etc. to copy these ideas so he stops missing them elsewhere. Filmed as Fable access was being banned mid-recording.

Standout features Theo wants every harness to steal:

Skills that run scripts

Claude Code skills can execute scripts when loaded (e.g. his repo-explorer skill pre-lists a clone cache before the model decides what to do), saving a whole agent step. He dismisses the "arbitrary execution" panic — it's no worse than npm and the code is auditable.

CLAUDE.md imports & overrides

`@path/import` syntax expands referenced files into context (recursive, 4 hops), so he points CLAUDE.md at AGENTS.md to unify the two competing standards without symlinks, plus Claude-only additions. `claude.local.md` (git-ignored) gives per-dev overrides without conflicts. Deep links (`claude://`) can launch sessions from HTML pages.

In-session UX

`/by-the-way` side chats without interrupting the main run; sub-agent navigation; rewind/branch; `/remote-control` to drive a desktop session from phone/web; flickerless full-screen renderer (`CLAUDE_NO_FLICKER=1`); proper git worktree support for ephemeral work.

Workflows + code mode

The killer feature: "Ultra Code"/workflows where Fable writes ~240 lines of throwaway JS to orchestrate staged sub-agents (audit → rule → verify) with dynamic prompts. Code as a step between model runs is more flexible than any built-in feature could be — but burns usage insanely (~$100/10 min; hit 24% of a window just reading the code). He also notes account-switching mid-run reroutes usage to a fresh sub on each new API call. Fable is shown to be excellent at prompting Anthropic models (it knows their failure modes) and, amusingly, hates GPT-authored code as much as Opus 4.8 loved it.