Claude Code daily workflow commands most people skip

This is the second of three reference posts built from Gui Ferreira’s NDC Oslo 2026 talk. The first post covered one-time setup. This one covers the biggest section of the talk, the commands and shortcuts you reach for during a normal working day.

🚀 Complete Claude Code & Coding Agents Course

Everything here is checked against the Claude Code commands reference as of November 2026. Version numbers are listed where the docs give them. If a command doesn’t show up for you, run /help and update Claude Code.

Planning

Edit the plan yourself with Ctrl+G

Press Shift+Tab to cycle permission modes until you reach plan mode, or type /plan followed by the task. Claude researches the code and proposes a plan without editing anything.

Here’s the habit Gui broke. When the plan is mostly right but one line is wrong, the instinct is to argue with Claude about it. “No, not that way. Change this part.” Each round costs a turn and adds noise to the context.

Press Ctrl+G instead. The plan opens as a markdown file in your editor. Fix the line, save, close, and Claude picks up your version and continues. Gui compares it to a shared Google Doc with Claude.

I compared plan mode with two other planning approaches in Plan Mode vs Grill Me vs Superpowers.

Make Claude interview you

Plans built from a short prompt are full of assumptions. You find out which ones were wrong weeks later, when someone writes the if statement nobody discussed.

Claude already asks multiple-choice questions sometimes, through its built-in question tool. You can force it:

Interview me about this feature. Don't stop until every open question is resolved.

It keeps asking until the plan has no gaps it can see. You make the decisions up front instead of finding them in the diff.

Undo and change direction

/rewind instead of “now refactor it back”

Say Claude asks whether you want MVC controllers or minimal APIs. You ask what it recommends, it says MVC, you agree. Twenty minutes later you look at the code and it’s too much for the problem.

The usual move is another prompt: “Refactor this to minimal APIs.” That works, but the whole MVC detour, the reasoning and the generated code, stays in the context window. Once the context is large, that history can drag down later answers.

Press Esc twice, or run /rewind, and pick the message to go back to. You can rewind the conversation, the code, or both, then rewrite the prompt and take the other path. Aliases are /checkpoint and /undo.

If you want to keep the current conversation and try another direction alongside it, /branch makes a copy you can switch back to with /resume.

Models and effort

Switch model and effort mid-session

A common pattern is a strong model for planning and a cheaper one for implementation. /model switches models mid-conversation. In the picker, press s on a row to switch for this session only instead of saving it as your default.

The part most people miss is effort. Each model can run at different effort levels, and higher effort means more reasoning tokens. /effort takes low, medium, high, xhigh, max, or ultracode, depending on the model. ultracode combines xhigh reasoning with automatic workflow orchestration. max and ultracode apply only to the current session. /effort auto resets to the model’s default.

So you can plan with Opus at high effort and implement with Opus at low effort. Your status line should show both, or you’ll forget which one is active.

ultrathink for one hard prompt

Changing effort for the whole session is easy to forget. Gui’s alternative is typing ultrathink in a single prompt. He says it runs that one prompt at max effort and then drops back to your previous level, and it triggers a rainbow animation. I didn’t find this keyword in the current commands reference, so test it in your version before relying on it.

Prompting while Claude works

Ctrl+S stashes your draft

While Claude runs a long task, you start writing the next prompt. Then Claude stops and asks you something. You used to copy your half-written prompt into a notepad.

Press Ctrl+S. Your draft is stashed and the input clears. Answer Claude, then press Ctrl+S again to bring the draft back. It also comes back automatically after you send another prompt.

/btw for side questions

If you type a normal question while Claude is working, it waits in the queue until the current turn finishes. And the answer lands in your main conversation.

/btw asks a quick side question without adding it to the conversation. You get an answer while the main task keeps running, and the main context stays clean.

/btw did we agree on repository pattern or direct Prisma calls?

Cost and context

/usage for what a session cost

Gui calls this one “status” in the talk. In current versions, /status shows version, model, account, and connectivity. Cost lives in /usage, with /cost and /stats as aliases.

/usage shows session cost, plan limits, and activity stats. On Pro, Max, Team, and Enterprise plans it breaks usage down by skill, subagent, plugin, and MCP server. That breakdown is useful because Claude starts subagents on other models without asking, and you’d otherwise never see what they cost.

/context in an empty session

/context shows context usage as a colored grid. Gui’s tip is to run it right after starting a session, before you type anything.

An empty session shows what loads without you asking for it: CLAUDE.md, rules, skills, MCP tools. If that’s already 8% of the window, something is too heavy. Skills each add a small footprint, and the ten skills you installed from someone’s LinkedIn post add up. Pass all to expand the per-item breakdown.

/compact before Claude starts flailing

There’s a popular idea that output quality drops once you pass about 40% of the context window. Gui calls it the “dumb zone” and is upfront that it’s opinion, not a study.

Claude compacts automatically near the limit. Gui’s advice is to do it yourself earlier. The signal is when Claude keeps guessing and failing on the same problem and you catch yourself typing “try harder.” Run /compact then. You can pass instructions about what to keep:

/compact keep the decisions about the data model and the failing test names

He also splits features into parts, like data model, then domain, then UI, and compacts between parts instead of starting a new conversation. The context took time to build, and a summary of it is still useful.

I go deeper on context in Claude Code Context Window: How to Stop Getting Dumber Answers.

/memory to audit what Claude remembered

Auto memory lets Claude write notes about you and your project. Sometimes it writes wrong ones. Gui’s story comes from a chat assistant’s memory. The talk names ChatGPT and Claude in the same anecdote, so I can’t tell you which. It kept calling him Emily because he once asked it to summarize a blog post that contained the sentence “my name is Emily.”

Run /memory to see your memory files and auto memory entries, and to turn auto memory on or off. Delete the wrong entries. Then look for the right ones. Claude sometimes picks up a real team preference from a single prompt. Move those into your project CLAUDE.md so the whole team gets them.

Shell and git

! runs a shell command directly

Start a prompt with ! and everything after it runs as a bash command. !npm test, !git status, !ls. No reasoning, no tokens spent on Claude figuring out how to run a command you already know.

The output still goes into the context window. Don’t use it on commands that print secrets.

/diff to watch changes

/diff opens an interactive diff viewer. Left and right arrows switch between the full uncommitted git diff and individual Claude turns. Up and down move through files. Since v2.1.198 the viewer also refreshes when git state changes outside the session, like a branch switch in another terminal.

Working away from the desk

/remote-control

Gui’s favorite, and the one he says changed how he works. /remote-control, alias /rc, makes the current local session available from claude.ai, including the Claude mobile app.

His example: he was behind on a delivery when his father called asking for a ride from the hospital. He started remote control, gave Claude a long task, drove to the hospital, and kept working from his phone in the waiting room. He also uses it to send Claude off on a multi-hour plan and go for a walk. Push notifications tell him when Claude needs approval or finishes.

It requires a claude.ai subscription.

/voice

/voice toggles voice dictation. You can pass hold, tap, or off to pick a mode. In hold mode, hold the space bar and talk.

Gui’s argument for it is rubber-ducking. Explaining a problem out loud to a colleague often solves it halfway through the sentence. He finds he understands what he’s building better when he says it. He also admits it’s strange in an open office. Requires a claude.ai account.

Paste screenshots

Screenshots work as input. Paste one of a site you like or a Figma frame, then describe what you want. Claude uses both the text and the image. You can drag and drop images into the terminal too. On a Mac, paste with Ctrl+V, not Cmd+V.

Running sessions in parallel

Worktrees, /color, and /rename

Git worktrees let you check out several branches of the same repo in separate folders. With Claude Code that means a feature, a code review, and a bug fix in the same project at the same time.

The problem is keeping track of which terminal is which. /color sets the prompt bar color for the session: red, blue, green, yellow, purple, orange, pink, or cyan. /rename names the session and shows the name on the prompt bar. Gui uses red for bug fixes, orange for refactors, blue for features, and names each session after its GitHub issue number.

/goal instead of babysitting

/goal sets a completion condition. Claude keeps working across turns until a separate check decides the condition is met. Added in v2.1.139.

A good goal includes how to verify it. Gui’s example is an invoice API:

/goal Build the invoice API. Done when an integration test creates an invoice,
calls the PDF endpoint, and the PDF matches the sample in fixtures/.

/goal with no argument shows status. /goal clear stops it early.

/focus to hide the noise

When you jump between several sessions, Claude’s running commentary isn’t useful. /focus toggles a view that shows only your last prompt, a one-line summary of tool calls with edit stats, and the final response. The setting persists between sessions.

It only works in fullscreen rendering. Turn that on with /tui fullscreen.

claude agents

To manage many sessions from the terminal, start with claude agents instead of claude. Agent view lists every session grouped by state: running, needs your input, finished. You can jump into any of them. Press the left arrow from any session to open it, and use /bg to send the current session to the background.

Agent view needs v2.1.139 or later and is a research preview. If you’d rather not live in the terminal, the Claude desktop app does the same job.

Quick reference

Command or key	What it does
Shift+Tab / /plan	Enter plan mode
Ctrl+G	Open the plan or prompt in your editor
Esc Esc / /rewind	Roll back conversation and code
/branch	Fork the conversation to try another direction
/model	Switch model. s in the picker for session only
/effort	Set reasoning effort level
Ctrl+S	Stash and restore the current prompt
/btw	Side question outside the main context
/usage	Session cost and plan usage. Aliases /cost, /stats
/context	Visual breakdown of context usage
/compact	Summarize the conversation, with optional focus
/memory	Edit memory files, review auto memory
!command	Run a shell command directly
/diff	Interactive diff viewer
/remote-control	Continue the session from claude.ai or mobile
/voice	Voice dictation
/color, /rename	Label sessions
/goal	Keep working until a condition is met
/focus	Show only prompt, summary, and result
claude agents	Dashboard for parallel sessions

The last post in this series covers extending Claude Code: hooks, skills, the bundled skills Anthropic ships, project MCP config, and headless mode.

Source

This post is based on Gui Ferreira’s talk at NDC Oslo 2026. The tips and anecdotes are his. The verification against the docs and any opinions are mine. Watch the full talk here: What Claude Code Can Do That You Haven’t Tried, Gui Ferreira, NDC Oslo 2026.