The autonomy ladder — six permission modes

Claude Code has six permission modes. Most people use one.

Approve everything and you're a doorbell. Approve nothing and you're gambling the repo. The ladder is how you stop doing either.

Autonomy scales with how recoverable the workspace is.

Not with how much you trust the model. The whole choice compresses into that one sentence.

The five rungs.

dontAsk sits off the ladder: it auto-denies whatever isn't whitelisted. CI and scripts, not humans.

• 01plan — reads only; proposes changes, makes none
• 02default — reads free; every edit and command prompts
• 03acceptEdits — edits and filesystem commands sail through
• 04auto — nearly everything, behind a safety classifier
• 05bypassPermissions — everything; disposable worktrees only

One keystroke, two flags.

# cycle modes inside a session
Shift+Tab    # default -> acceptEdits -> plan -> ...

# start a session on a rung
$ claude --permission-mode plan

# unlock rung 5 - spelled like a warning label
$ claude --dangerously-skip-permissions

The studio runs that last flag every working day — and almost never types it. It's baked into the dev script, where a worktree contains the blast radius.

Trust is not a safety strategy.

Plan to look. Default to decide. Bypass to burn.

plan to look, default to decide, acceptEdits to iterate, auto to flow, bypass to burn — but only what you can afford to burn.

Match the rung to the workspace.

• Client repo, no remote — plan or default; nothing is recoverable yet
• Tracked repo, reviewing diffs after — acceptEdits
• Daily driver, backed up — auto, the studio's resting default
• Disposable worktree — bypassPermissions
• CI and scripts — dontAsk, pre-approved rules only

One idea from one lesson.

Track 1 (23 lessons) is free.