The /config Cheat-Sheet: Change Any Claude Code Setting Mid-Session

Until recently, changing a Claude Code setting meant the same annoying ritual: quit your session, open settings.json, edit the value, save, and start over — losing your place mid-task. Claude Code v2.1.181 (released 2026-06-17) killed that ritual with one tiny command. Type /config key=value right in the prompt and it sets any setting on the spot — no exit, no file edit, no restart. The changelog's own example is /config thinking=false: turn off extended thinking for a quick, throwaway task, then flip it back on with /config thinking=true when you want depth again. You're steering the model live instead of babysitting a config file. It works everywhere Claude Code runs: the normal interactive chat, headless runs with claude -p, and Remote Control sessions.

The shape is dead simple: a slash, the word config, then a key=value pair with no spaces around the =. `` /config thinking=false ` That applies immediately, in the same session you're already working in. Before v2.1.181, the equivalent meant editing settings.json and relaunching — which dropped your context and your train of thought. Now the change is a single line you can fire between prompts. The headline word in the changelog is any: /config sets any setting from the prompt, not a hand-picked few. The most common one to flip live is thinking`, but the mechanism is general — the value you pass is whatever that setting accepts (here, a boolean).

Extended thinking is brilliant when you want depth — debugging a gnarly race condition, planning a refactor. It's overkill when you just want a one-liner renamed or a quick fact. Paying the latency of deep reasoning on a throwaway task is wasted time. So: drop into a fast mode for the quick stuff, then climb back to depth when it matters — without ever leaving the session. - /config thinking=false — extended thinking off. Claude answers fast instead of pausing to reason it all out. Great for quick edits, renames, lookups, and 'just do the obvious thing' tasks. - /config thinking=true — extended thinking back on. Deep reasoning returns for the hard problems. The point isn't the one setting — it's that you can now tune the model's behaviour as you work, matching the tool to the task in real time.

/config isn't limited to the interactive REPL. Per the changelog it works in three places, which is what makes it genuinely useful for automation as well as hands-on coding: - Interactive chat — the normal Claude Code session. - Headless -p — claude -p non-interactive runs (scripts, CI, agents). You can shape behaviour per-invocation without touching shared config. - Remote Control — Remote Control sessions honour it too. That last point matters: because the same command works in headless and remote contexts, you can flip a setting for a single automated run without permanently changing your settings.json for every other session.

The whole trick collapses to a single habit worth memorising: when you want to change how Claude Code behaves right now, don't reach for the settings file — type /config key=value. The one to internalise first is the thinking toggle, because you'll use it daily: `` /config thinking=false # quick task, answer fast /config thinking=true # hard problem, reason deeply `` Keep this page next to your terminal. As you discover other settings worth flipping on the fly, the syntax never changes — only the key and the value do.