Hidden Claude Code Settings That Change How You Work

0:00Claude Code is packed with settings, commands, and keyboard shortcuts that most people — even daily users — never touch. This video is a curated tour of the ones the channel says "changed how we work," from a flag that fixes /init to a command that lets separate sessions talk to each other. It's a dense listicle, so this deep dive groups them by what they actually solve, normalizes the auto-captioned command names, and flags what's verified vs. described-from-video. (There's a sponsor read for Scribba's "Explain" tool mid-way; it's noted and skipped here.)

Setup that learns how you work

0:22The first tip is the one the video claims the Claude Code team themselves use: a fix for /init. By default, /init reads your project and dumps everything it finds into CLAUDE.md — but that file is supposed to hold only what the agent must always remember, not things it can re-derive by reading the code. The fix is a flag (the video renders its name unclearly; it's set before launching claude) that makes /init stop auto-generating and instead ask you questions: improve the current setup or start fresh, generate skills and hooks, how you use Git, how you'll deploy. From your answers it drafts a small plan, waits for approval, then writes a CLAUDE.md that reflects your actual way of working — plus starter skills and hooks.

1:57If you're coming from Codex or Cursor, /import migrates your existing setup instead of rebuilding from zero. Run it with the name of your previous agent and Claude Code lists what it can copy — instruction files, connected MCP servers, and the rest — then you pick. It wraps the selection in a skill and uses that skill to copy everything into Claude Code.

Keeping setup from going stale

2:34Instructions written for older models go bad as models improve, and they end up wasting context rather than helping. The video's fix is an audit command: run it and Claude Code reads your whole setup, uses subagents to inspect each part, and returns a report of every problem plus suggested fixes. Apply them and only the instructions that still help survive. The video flags this as most useful right after a new model ships, when the most instructions go stale at once.

3:28Skills rot too — most sit in context unused. skill-doctor shows how often each skill is used, when it was last used, and how many tokens it occupies in the context window, so you can delete the dead weight. The companion /skills command lists every skill, and pressing T sorts that list by token count ascending, so you can see exactly what's bloating your context.

Permission tuning

4:15Auto mode lets safe commands through without asking, but it doesn't know your habits. The auto-mode setup command reads your project and past session history, then rewrites auto mode's rules to match the commands you routinely allow — showing you the new setup for approval first.

4:52If you'd rather not touch auto mode's config, there's a "fewer permission prompts" command that mines your past sessions for the read-only commands and tools you use most, adds them to an allowlist, and saves it to the project's settings file — so long unattended runs can proceed without stopping for prompts, but can't change anything you didn't already sanction.

Freeing context

5:24Two settings reclaim context that's consumed before you even type. disableClaudeAIConnectors (in settings.json) stops the Claude-app connectors — Notion, Slack, and friends — from loading into every Claude Code session when you never use them. And 6:06enableArtifacts: false turns off the artifacts tool, which also loads every session and uses context if you don't need design previews. Both live in the same settings file; you can ask Claude Code to edit them for you rather than opening the file yourself.

Evals: build-eval & hillclimb

6:29If the app you're building uses AI, you need to test whether its AI gives good answers. The claude-api skill ships two workflows (added in late Sept 2026). build-eval generates a reusable test suite — written once, so every new version of your app is checked against the same tests. Then 6:52hillclimb iterates: it makes a change, runs the tests, scores the response, and keeps only the changes that score higher than the previous one. Both cost real API money per run because each test calls the model — and if your app doesn't call a model at all, build-eval still helps by planning tests for you without the API.

Plan mode & output

7:24Plan mode's "clear context before building" option quietly disappeared. The setting showClearContextOnPlanAccept: true brings it back, so you can drop all the planning back-and-forth and start building with only the plan in context. If you do, the plan isn't lost — it's saved to a plans folder under ~/.claude. 8:04To keep plans in your project instead (so you can share them), set plansDirectory to a project folder.

9:37The /output-style command switches among five styles: default, proactive (act first, plan later), concise (shorter answers), explanatory (explains more of what it's doing), and learning (asks you to write some of the work so you learn while building). 10:14And showThinkingSummaries: true restores the Ctrl+O reasoning summaries that used to be shown but no longer are.

Session management

10:37Focus mode (/tui set to full-screen, then /focus) shows only your prompt and Claude's answer plus a short summary — the intermediate steps scroll off when done, so you don't hunt through a long task when you return. 11:17/branch copies the current chat into a new session so you can chase an idea without losing your place; 11:32/fork is branch but the copy starts in the background with a task you hand it, ending itself when done. 11:51And /subtask starts a subagent with your full context — unlike a normal subagent, which gets a fresh context and only the prompt it's given.

12:20The flagship feature here: sessions can now talk to each other. list-agents shows which sessions are reachable, and you can ask one session to message another that already worked out the answer — receiving the answer even while the second session is still working. 13:01Two settings round out multi-session work: askUserQuestionTimeout (in seconds) makes Claude pick the suggested answer and keep going when you're away, though it only self-answers when it can proceed without you; and 13:37/config → local notifications (e.g. a terminal bell) tells you when a session needs your input or finished.

Model routing

14:00/model opus-plan makes plan mode run on Opus while everything else builds on Sonnet — the stronger model for planning, the cheaper one for execution, which matters most on the Pro plan where limits hit fast. 14:26For subagents, set CLAUDE_CODE_SUBAGENT_MODEL to a smaller model and CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 so a subagent's own file can't override it (skip the second flag if you want the file to win). Both are read at session start.

Keyboard shortcuts & /powerup

15:06Three prompt-editing shortcuts: Ctrl+S stashes the current prompt (empties the box, restores it later), Ctrl+Y restores deleted text, and Ctrl+_ undoes the last change. Finally, 15:42/powerup runs short interactive lessons inside Claude Code, each an animated demo of a feature so you learn as you work.

Claims checked

Sources & further reading