Mariana's Claude Code Tips

Paste this into your next Claude Code session

For any substantial build, use this two-step instead of just asking for the thing:

Prerequisite: the superpowers plugin, a free add-on pack that teaches Claude this way of working. Don't have it? Just ask: "install the superpowers plugin". Claude handles it.

What it does: Claude first writes a detailed implementation plan (every task with exact files, values, and test steps), then executes it by dispatching a fresh sub-agent per task. Each one gets a written brief, builds, and self-reviews; then a separate reviewer that never saw the work happen tries to poke holes in it before the next task starts. The main session only coordinates; it never writes or reviews the code itself. Fresh eyes catch what the author can't.

Why it saves money too: the smartest model does the planning and checking, while cheaper, faster agents do the labor. Best of both.

Make it your standing default

Never remember the phrase again

Tell Claude Code:

Make this the standing default for big builds: add it to my CLAUDE.md.

Claude adds the rule to your CLAUDE.md (the memory file where Claude keeps your standing instructions), and every future build big enough to qualify uses the two-step automatically. You can always opt out per build ("quick and dirty, inline").

📖 Official docs: Plugins ↗

Paste this into Claude Code: it installs itself

One paste and Claude does the setup for you. It installs the starter kit's statusline (the strip of info along the bottom of Claude Code), and the tab names come with it:

Needs a reasonably current Claude Code and a small helper tool called jq. If jq is missing, Claude will tell you and help you install it.

Or grab the whole starter kit

Two commands, full setup

The kit was built to be the answer to "how do you use Claude Code?": statusline, sounds, and more, installed as a plugin:

/plugin marketplace add https://github.com/marianasmall/claude-code-starter-kit
/plugin install kit@claude-code-starter-kit

Installing as a plugin means updates are one command away: when new goodies land in the kit, run claude plugin update kit@claude-code-starter-kit in your terminal (not inside Claude Code), restart Claude Code, then run /kit:verify, and you're current. Not sure which plugins and connectors are worth adding? The kit's plugins and connectors primer walks through it.

Related: Fun 03, the mascot status line, lives in the same bar. The kit's statusline and the mascot are cousins; you can merge them.

📖 Official docs: Status line ↗

Right after finishing a job you'll do again, paste this

While the work is still fresh in the conversation:

The ladder: a prompt you repeat word-for-word becomes a slash command (Fun 07). A whole procedure you repeat, with steps, checks, and judgment calls, becomes a skill. Claude uses it automatically when the job matches, or you call it by name.

How to know it's skill-worthy: you've done it two or three times, it has steps you'd forget, and you keep correcting Claude the same way each time. That correction is exactly what belongs in the skill.

What a skill looks like inside

Just a folder with instructions

A skill lives at ~/.claude/skills/[name]/SKILL.md, written in plain markdown:

---
name: monthly-report
description: Build the monthly client report from the analytics exports
---

1. Ask which month we're reporting on
2. Pull the exports from the reports folder
3. Follow last month's structure: sections, tone, charts
4. Flag any metric that moved more than 20% before writing
5. Save the draft as Report_[Month]_[Year].md and open it

You never have to write this by hand; that's what the paste prompt above is for. Claude drafts it from the work you just did together; you just approve it.

📖 Official docs: Skills ↗

Let Claude interview you

Paste this into Claude Code and answer its questions:

The file lives at ~/.claude/CLAUDE.md (applies everywhere) or inside a project folder (applies just there). It's plain text: read it, edit it, delete lines that stop being true.

The compounding trick: whenever you correct Claude the same way twice, say "add that to my CLAUDE.md." The file gets smarter every week; your sessions stop starting from zero.

What actually belongs in it

Four kinds of lines earn their place

• Who you are: "Marketing exec, not a coder. Explain what things accomplish, not the syntax."
• Standing rules: "Always back up a file before overwriting it." "Explain before acting."
• Preferences: output format, tone, how much detail you want.
• Corrections that keep recurring: anything you've typed twice belongs here, not in chat.

Want a filled-in starting shape instead of a blank page? The starter kit ships a CLAUDE.md.template. And the "standing default" upgrade in P1 is exactly this: one line in your memory file.

📖 Official docs: Memory ↗

Paste this at the start of any real project

No installs needed. Claude sets up the files and the habit:

The payoff moment: next time a session ends (or dies), you open handoff.md, copy the bottom half, paste it into a fresh thread, and Claude continues like nothing happened.

Why five files, and what each one does

Each file answers one question

• README.md: what is this? (for humans)
• PLANNING.md: where is it going? (goals, roadmap)
• CONTEXT-SUMMARY.md: what's true right now? (updated each session)
• CLAUDE.md: project-specific rules for Claude (only if needed)
• handoff.md: where did we leave off? (replaced every session)

The handoff's Traps line is the sleeper hit: dead ends already tried, decisions already made. A new thread that knows the traps doesn't waste your evening re-hitting them.

📖 Full conventions + templates: project-conventions.md ↗

Just ask Claude to go look

Paste this with your topic:

Keep them longer: 30 days is the default, not the limit. One line in ~/.claude/settings.json:

{ "cleanupPeriodDays": 120 }

Sessions are just files

Every conversation lands in ~/.claude/projects/ as a transcript file. Nothing leaves your machine. That's why Claude can read them and why the retention setting is yours to control.

Related built-ins: /resume reopens a past session where it left off, and /recap summarizes what happened when you return to one. The starter kit ships /kit:recall, a ready-made version of the search paste above.

📖 Official docs: Settings (cleanupPeriodDays) ↗

Ask Claude to design your setup

Paste this into any session:

The three rules that make it work: shared files beat memory (sessions coordinate through what's written down, not what anyone "remembers") · update the handoff when state changes, not just at the end · when work moves to a new session, the old one writes the new one's first message.

Going deeper: same repo, two sessions

Worktrees: parallel work with no collisions

Two sessions editing the same project will fight. Git worktrees give each session its own copy of the repo on its own branch: same history, separate desks. Ask Claude: "set up a worktree so I can work on [thing] in a second session without touching this one."

📖 The full primer: how sessions "see" each other ↗

Four habits, each one a single command

Anything that starts with a slash is just something you type into Claude Code and press enter. It isn't code, and there's nothing to install. You don't need all four habits at once; pick the one that matches your day.

1. Starting something unrelated? Type /clear. It begins a fresh conversation. Without it, the invoice you finished is still in the conversation while you ask about a press release, and Claude keeps re-reading it all afternoon. If you'll want the old conversation back, type /rename first to name it, then clear. Reopen it later with /resume.
2. Sitting down for a long session? Type /context. It shows what's already in the conversation before you've typed a word, and how much room each thing takes. Every connected tool (Gmail, Drive, a spreadsheet app) describes itself at the start of every message, whether you use it that day or not. Type /mcp to switch off the ones you don't need. (MCP is just the name for those connections.)
3. Pick your model at the start, then leave it alone. Switching with /model halfway through makes the next reply slower, because the new model has to read everything from the beginning. Turning on fast mode mid-session does the same. Bigger isn't automatically better: a smaller model is quicker and handles precise, well-described or mechanical jobs well. Keep the biggest for work that's genuinely hard or vague.
4. Tidy up before you walk away, not after you come back. On a Claude subscription, Claude keeps your conversation quick to reload for about an hour. Heading into a meeting, or stopping for the day? Tell Claude what to keep and compact first:

That swaps the long conversation for a short summary that keeps what you named. Do it an hour later and Claude pays full price for one last complete read first.

Claude has no memory between messages

Picture a folder of notes on a desk. Every time you send a message, Claude is handed the whole folder and reads it from page one: its own instructions, every document it has opened, everything you've both said. Only then does it reach your new message at the very end. Ask forty questions in one sitting and on the fortieth it's re-reading the previous thirty-nine.

One thing helps. Pages at the front of the folder that are exactly the same as last time get skimmed, which is much quicker. But change something near the front and Claude has to read properly again from that point on. The four habits are all ways of keeping the folder thin, or keeping the front of it still.

If you're on a subscription, "cheaper" means "you get further before Claude makes you stop." Plans have a usage window. Work hard for long enough and Claude tells you to wait. These habits push that point back. The skimmed copy expires if you step away for a while, so the first message after a long break is the slow, expensive one.

Why /clear beats /compact for a fresh start: compacting has to read the whole conversation to summarize it, so it isn't free. Clearing is. Use /compact when you want to keep the thread, and /clear when you don't.

📖 Official docs: Manage costs ↗ · Anthropic's write-up: maximizing the value of your sessions ↗

Three smaller ones

• Tell Claude which file you mean. "Fix the pricing in the September deck" sends it hunting. Type @ and pick the file instead, and the hunt is skipped.
• Heading the wrong way? Type /rewind. It tears out the last few turns and takes you back to an earlier point. Unlike compacting, it keeps what came before intact, so it costs almost nothing. Better than arguing with it.
• Push messy jobs to a sub-agent. A sub-agent is a second Claude working in its own separate conversation. It reads the messy thing (a full meeting transcript, a long list of files) and comes back with just the answer. Say "do that in a sub-agent."

When Claude gives you something wrong

Ask which kind of wrong it was, because the fix differs. If it didn't know enough (went the wrong way, missed how things fit together), move up to a bigger model with /model. If it didn't try hard enough (skipped a file, stopped halfway, went shallow), raise its effort with /effort. It works the other way too: if Claude writes an essay when you asked it to change one word, turn the effort down.

Write the instruction once

Tell Claude what to keep every time it compacts

Claude also compacts on its own when a conversation gets full. Paste this so it knows what matters when that happens:

CLAUDE.md is the memory file from P4. The new section takes effect in your next conversation.

Make it yours: this is how one person runs it. Tell Claude which of these habits you want, and it will set them up to suit how you work.

📖 The full primer: context budget ↗

Connect your recorder, then ask

1. Check the recorder is signed in to your own account, with cloud sync on. Recordings land in whichever account the device is paired to, and Claude reads the cloud copy, not the device. If you inherited a device from someone else, they need to unpair it first.
2. In claude.ai, open the left sidebar and go to Customize > Connectors. Search for your meeting recorder or note-taking app. If it's listed, click Connect and sign in with the account from step 1.
3. Verify it. This step isn't optional. Paste this:

Then ask: "Summarize the newest one and list the action items." If you see your latest recordings, you're set. If the list is empty, the device is paired to a different account, or its cloud sync is off.

4. Make the habit that pays off. While you record, say instructions out loud with a fixed prefix, like "for the bots: remind me to send the deck." Later, ask Claude to find everything you said for the bots in the last few days. Those lines are short and explicit, so they're easy to miss in a long transcript and easy to find by the phrase.

Then ask the questions you actually have: "What did I promise in my last call with a client?" or "Which recordings mention the budget?"

Each person connects their own account. Connectors are switched on per account, and everyone signs in with their own login. On Team and Enterprise plans, an admin may need to switch a connector on before members can connect it. Not every recorder has a connector. If yours isn't in the list, this tip won't work yet.

What it can do: find recordings by name, keyword, or date; read the full transcript; read the app's own summary and action items. It reads. It can't delete or change anything.

One honest caveat: what Claude reads is the recorder's cleaned-up version of the transcript. It's good for summaries and action items, not for quoting someone word for word. Names and unusual words often come out wrong, especially in noisy rooms. Check any name before you put it in an email, and listen to the original when exact words matter.

Think about what gets recorded. A recorder that's always on picks up everything, including calls you wouldn't want anywhere else. Connect only the account you're comfortable with Claude reading.

It works in Claude Code too

Nothing extra to install

The connection follows your Claude account. It works in the desktop app, on your phone, and in Claude Code. In Claude Code, /mcp shows your connected tools once they've carried over. Not sure which connectors are worth adding? The kit's plugins and connectors primer walks through it.

Make it yours: this is one way to run it. Treat it as a starting point and tell Claude what to change, like which phrase you use or what you want pulled out of each meeting.

Connect your calendar and inbox, then paste

1. Connect Google Calendar and Gmail. In claude.ai, that's Customize > Connectors. In Claude Code, type /mcp. You sign in once in a browser window and approve access.
2. Paste this:

A good first brief names real meetings and real senders. If it's vague, say so. If it misses an email you'd expect, tell it what to always include, like "anything from my accountant." Claude will use that next time.

Why it works: the brief isn't clever. It saves the few minutes you'd spend hopping between your calendar and inbox before you start. Because it only reads, there's nothing to undo.

If you travel, say which time zone to use. If you keep work and personal calendars apart, say which ones to check. Claude reads your main calendar unless you tell it otherwise.

Make it a habit

Three ways, from most hands-on to least

• Keep it manual. Each morning, say "run my morning brief."
• Make it a command. Ask Claude to save the prompt as a slash command called /morning-brief, then type that each morning. See Fun 07.
• Schedule it. Ask Claude to set it up to run every weekday morning with the /schedule command. It then runs without you opening anything.

Want it somewhere you'll actually see it? Ask Claude to put each brief in a draft email to yourself instead of the chat.

📖 The full recipe, with variations ↗

Add one standing rule

What you get: events created as the work happens, not written down afterwards. Each one carries the real locations: file paths, which folder, which project, the web links, and which account to sign in as. Where a login is needed, it says where the password lives, never the password itself. And for work you'll carry on later, there's a block you can copy on your phone and paste into a fresh Claude chat to restart exactly where you left off. A half-finished job becomes one copy and paste away from restarting instead of a re-explanation.

Step 1: connect your calendar so Claude can create events. In claude.ai, that's the Google Calendar connector under Customize > Connectors. In Claude Code, you need a calendar connection (Google Calendar or Outlook); /mcp shows what you have. Without it Claude can't create anything.

Step 2: install the rule. It has to live somewhere Claude reads every time, or it works once and disappears. In Claude Code, paste this and it goes into your memory file:

CLAUDE.md is your memory file (P4). In claude.ai, paste the quoted rule into your custom instructions in Settings instead, or into a Project's instructions to limit it to one body of work. Pasted into a single chat, it applies to that chat only.

Then check it worked. Do this once, on purpose. End a real session with something that's blocked on you, then open the event on your phone. Can you do the thing from what's written there, without opening your laptop and without asking a follow-up question? If not, tell Claude what was missing. It calibrates fast, and that correction is worth more than any wording you could have specified up front.

Locations are the difference between a reminder and a working instruction. An event that says "send the deck" without saying where the deck is just sends you back to Claude to ask. That's the interruption the event was supposed to replace.

Only things waiting on you belong there. If everything becomes an event, the calendar stops meaning anything. Work Claude can do itself should just get done.

One honest caveat: the reliable half is capturing each item at the moment Claude hits something only you can do. The "sweep back at the end of the session" half is weaker, because Claude has no way to tell that a session is ending. If you want that sweep, hook it to something you actually do: a wrap-up command you run, or a phrase you say. Otherwise it will quietly never fire.

Work you'll pick back up later

A paste-ready block, inside the event

When the next step is "keep going on this project," ask Claude to end the event description with a block you can paste into a fresh conversation: what the work is, where everything lives, what's done, what's next, and what not to touch. Your phone is then enough to restart the work.

Keep work and personal calendars apart? Say which is which in the rule, or Claude will use your main one.

Or let Claude do the setup

Prefer one line? Paste this and Claude reads the guide and sets the rule up for you:

Make it yours: this is how one person runs it, tuned to how they work. Treat it as a starting point and tell Claude what to change: the timing, the level of detail, or what deserves an event at all.

📖 The full primer: calendar handoff ↗

Paste this with a name and a project

Works best with your email and calendar connected (see P14). Without them, Claude will ask you to paste in the key threads.

What you get back:

• A timeline of who was waiting on whom at each step
• A who's-who, and open items with owners and ages
• Every claim tagged by source strength: an email beats a transcript, which beats a spreadsheet
• An honest "what the record can't tell us" section instead of confident guessing

Read the brief yourself before it goes to anyone. Claude shows you what it found and what it left out, so you decide what the reader sees.

One honest caveat: it can only read what your Claude can reach, and it's built to tell you what it couldn't see rather than papering over gaps. Big inboxes take a few minutes.

Not all sources are equal, and the brief says so. Original emails and documents count most. Meeting transcripts are useful but machine-cleaned, so they're never treated as exact quotes. Anyone's memory counts least. A claim that rests on a transcript or a recollection gets checked against something stronger or marked as unverified.

It names the blind spot. The gap between when the record starts and when the reader first got access is where most confusion lives. The brief marks it instead of guessing.

Not using the kit? Install just this skill

In Claude Code (the terminal app, not the claude.ai website), paste this:

The skill can also draft a call agenda and a ready-to-send email to whoever owes answers. Say so in your ask. Claude drafts the email for you to review. It doesn't send it.

Make it yours: treat the skill as a starting point. Tell Claude what to change: which tools to check first, how long the brief should be, or what to leave out.

📖 Read the skill ↗

Connect, sign in once, then ask

1. In claude.ai, open the left sidebar and go to Customize > Connectors.
2. Find Fireflies, Wispr Flow, or both, and click Connect. You sign in once and you're done.
3. Check it worked. Paste this:

If you use Wispr Flow, also try "find my note about the benefits call." If the answer comes from a real meeting or note of yours, you're set.

On a Team plan: an admin can add a connector for the whole organization, but each person still authorizes their own account. Your transcripts stay yours until you connect them.

Your memory of a call drifts. The transcript doesn't. Asking Claude means the answer comes from what was actually said, with the date of the meeting it came from.

One honest caveat: transcripts are machine-made. Names and unusual words often come out wrong, so check a name before you put it in an email. Same advice as P13, which covers recorders in general.

Running Claude for a team?

Give every setup change the same subject line

Any time you change your team's Claude setup (connectors, plugins, skills, policies), announce it in an email whose subject starts with [Claude Update]. Anyone can search that phrase in their inbox and get the full history in one place.

Make it yours: pick whichever tag your team will remember. What matters is that it never changes.

Paste one line, then follow along

This needs Claude Code (Claude's terminal app) on a Mac or Linux. It doesn't work on the claude.ai website, which can't run it.

1. Open Claude Code and paste this:

2. Follow its checkpoints. The guide walks you through the Google console clicks one checkpoint at a time and waits for you at each one. Your part is about 10 minutes of clicking to create your own credentials, plus one approval per account.
3. When it asks which accounts and what to call them, give each a short name, like work, personal, or a brand name. Short names are what you'll type later.
4. Next morning, say "sweep my inboxes." Correct it once or twice ("replies to this client go from work," "always flag anything from my accountant") and it sticks.

The one step people skip: the guide has you publish your Google app to "production." Skip it and Google logs you out every seven days. The guide flags it. Take it seriously.

What you get: Gmail, Calendar and Contacts for every account you connect. One question covers all of them.

It cannot send and it cannot delete. Drafts only; you press send. Anything it wants gone gets a "To Be Deleted" label for you to confirm. Every action it takes is written to a local log.

One honest caveat: the no-send, no-delete guardrails are on by default because I wanted them. If you'd rather yours send, tell your Claude to change it. It's your build.

Google Drive works a different way

No connector needed

This setup covers Gmail, Calendar and Contacts. For Drive, install Google Drive for desktop and sign in to every account. Each one appears on your Mac as its own folder, and Claude Code reads them like any other folder, side by side.

Make it yours: this is how one person runs it, tuned to how they work. Treat it as a starting point and tell your Claude what to change (accounts, the sweep format, level of detail). It will rebuild any of it to fit how you think.

📖 The open-source connector and its setup guide ↗

Pick one and sign up

Online workshops (Anthropic and Tenex)

• Claude Workshop (Cowork): research, drafts, decks and follow-ups from your real work. No coding.
• Claude Code Workshop: for people building with Claude Code. Dates come and go, so check the events page for the next one.

1. Open the Anthropic events page and filter to Virtual.
2. Click Apply on a date and fill in the short form. The join link arrives by email.

In-person meetups (Claude Community)

Free, community-run, all over the world: luma.com/claudecommunity. Popular ones fill up and go to a waitlist, so register early.

Why bother: I took the Claude Code session on October 2. In 90 minutes it changed how I set up my projects, how I keep Claude focused, and how I plan before I ask it to build. It's hands-on: you build along with the instructor, and you keep the materials for 30 days.

Three habits you can use right away

1. Give Claude a briefing it reads every time (Project instructions in Claude, a CLAUDE.md file in Claude Code, see P4), including a map of where your things live.
2. One chat per task. Long chats cost more and drift. P12 has the details.
3. For anything important, get a second opinion. Paste the result into a fresh chat and ask it to find the problems:

Make it yours: this is how one person runs it. Tell your Claude what to change and it will adapt any of this to fit how you think.

Press Shift+Tab, or just ask for it

Tap Shift+Tab until the plan-mode indicator shows (or type /plan). No setup at all. And this paste works even if you forget the keystroke:

When it earns its keep: anything scary: deleting, reorganizing, touching many files, changing something that works. You read the plan in plain English, adjust it, then let Claude run.

The modes, in plain terms

Shift+Tab cycles between them

• Auto: Claude runs routine actions itself and checks with you before risky ones. The default for everyone since Claude Code 2.1.284.
• Manual: Claude asks before each action. For when you want to watch every step.
• Plan first: Claude designs the whole approach, you approve once, then it executes.
• Accept edits: Claude edits files and runs basic file commands without asking, and still checks with you before other commands. For changes you'll review afterward.

The status line under the prompt shows which mode you're in. To pick your own default, set permissions.defaultMode in ~/.claude/settings.json. Using the starter kit? Its safety hooks run underneath every mode, auto included.

The habit that sticks: plan mode for anything irreversible, auto for the everyday, manual when you want to see each step.

📖 Official docs: Permission modes ↗

Start with an audit: Claude finds what's exposed

Paste this into Claude Code:

That last sentence matters: a security audit that prints your keys into the chat has just copied them somewhere new. Make "never print values" part of the ask.

Keys served at runtime, not stored in files

1Password's Environments feature (desktop app) serves your keys through a secure pipe that scripts read the moment they need them, while your vault is unlocked, and never from a file on disk. Other password managers have similar CLI patterns; ask Claude which fits yours.

Belt and suspenders: also ask Claude to install gitleaks, a pre-commit scanner that blocks any commit containing something that looks like a key. Plaintext leaks mostly happen via git; this closes that door.

📖 The full primer: 1Password Environments ↗

Four standing rules, one paste

This goes straight into your memory file (P4):

Why this changes everything: once nothing can go permanently wrong, you stop supervising every step, and that's when Claude becomes genuinely useful. Reversibility is what trust is made of.

Reversible > confirmed

Asking "are you sure?" before every action is exhausting and you'll start clicking yes blindly. The better design: make actions undoable instead of confirmed. A review folder you empty weekly beats a hundred "are you sure?" prompts, and a draft you read before sending beats an apology email.

Pairs naturally with P6 (plan mode): see the plan before it runs, and even if something slips through, it's reversible anyway. Two layers, no fear.

📖 The starter kit's safety hooks ↗

A contacts list becomes the permission

First, put your work contacts in one place: a list in the Mac's Contacts app (File > New List; Apple's how-to; older Macs call it a group), or a label in your work Google contacts. Contacts shows iCloud and Google side by side, so either works. Then paste this into Claude Code:

Expect one Mac permission step along the way (Full Disk Access for your terminal app); the guide walks you through it.

Then verify. This step is not optional: ask for a work thread (it should come back), then ask for a real personal contact (it should come back empty). Have it draft to a fake number to prove nothing can deliver. A test only proves what it was written to prove, so use a real personal contact, not a made-up one.

The Mac can't scope this for you

macOS permission for reading Messages is all or nothing; there's no OS setting for "work texts only." The scoping lives in how the reader is built, which is why "cannot return anything outside that list" is the load-bearing phrase in the prompt, and why the verification step isn't optional.

The principle: make the scope structural, not a filter. Personal messages aren't filtered out of the results; they're in data the tool never opens. There's no filter to get wrong.

The payoff: the list is the permission. Add someone in Contacts and Claude can see that conversation; remove them and access is gone. No settings to edit, no list maintained in two places. The one hard rule: it never sends. Drafting is fine (it opens Messages with the reply loaded and you press send). Because no send capability exists in the code, it can't text anyone even if asked.

Not on a Mac? The texts part needs one: iMessage's database only exists on Apple devices. But the trick isn't Mac-only: the same work-contacts label can scope Claude's email or calendar reading instead. Same wording, same verification step.

📖 The full primer: scoped iMessage access ↗

Ask Claude.ai to write it for you

Paste this into Claude.ai and swap in a theme you love:

Themes that work:

• ADHD humor ("Starting five things at once")
• Portuguese / Brazilian slang ("Na boa", "Só um segundo")
• Travel ("Upgrading (manifesting)", "Already at the gate")
• Your profession (marketing, law, parenting, wine)
• Pop culture (Taylor Swift lyrics, Star Wars, The Office)

Generate a few batches with different themes and merge them. That's how power users end up with 150+ spinners.

Rather not edit the settings file yourself? Paste the same prompt into Claude Code instead and add: "then add it to my settings for me, and show me the change first." Or use the walk-through prompt at the top of the page.

Or write it yourself

Where it goes

Open ~/.claude/settings.json (create it if it doesn't exist) and add:

{
  "spinnerVerbs": {
    "mode": "append",
    "verbs": [
      "Pondering",
      "Scheming",
      "Caffeinating",
      "Reading the room",
      "Not starting with 'Great question!'"
    ]
  }
}

"append" keeps Claude's built-ins and adds yours. "replace" uses only your verbs.

📖 Official docs: Settings reference ↗

Ask Claude.ai to write them for you

Paste into Claude.ai with the theme you want:

Themes that work:

• Pep talk for ADHD brains
• Reminders a parent would give you
• Quotes from your favorite author
• Zen koans / short philosophy
• Inside jokes with your team

Rather not edit the settings file yourself? Paste the same prompt into Claude Code instead and add: "then add it to my settings for me, and show me the change first." Or use the walk-through prompt at the top of the page.

Or write it yourself

Where it goes

Add this block to ~/.claude/settings.json:

{
  "spinnerTipsOverride": {
    "tips": [
      "This is fine.",
      "Trust the process.",
      "You got this.",
      "Shipping beats perfect."
    ],
    "excludeDefault": false
  }
}

excludeDefault: false mixes yours with Claude's built-in tips. Set to true to use only yours.

📖 Official docs: Settings reference ↗

Ask Claude.ai to write the script

Never written a script? You don't need to. Describe what you want and Claude.ai will write the script:

Examples to ask for:

• "Current git branch, today's date, and a rocket emoji"
• "A motivational quote that changes every time it refreshes"
• "Whether my internet is working plus battery percentage"
• "The project folder I'm in and time of day (morning/evening)"

Rather skip the saving and wiring? Paste the same prompt into Claude Code instead and add: "then save the script and set it up for me, and show me each change first." Or use the walk-through prompt at the top of the page.

Or write it yourself

Simple version: static mascot

Put this in ~/.claude/settings.json:

{
  "statusLine": {
    "type": "command",
    "command": "echo \"🤖 Claude Code, let's build\""
  }
}

Dynamic version: live script

Point to a shell script that refreshes:

{
  "statusLine": {
    "type": "command",
    "command": "~/.claude/statusline.sh",
    "refreshInterval": 10
  }
}

Then create ~/.claude/statusline.sh and run chmod +x ~/.claude/statusline.sh to make it executable.

📖 Official docs: Status line ↗

If you're reading this and already seeing ★ Insight boxes or running commentary in your own Claude Code, you already have an output style enabled (Explanatory). This section is how to switch it up.

Ask Claude.ai to write the style file

Describe the persona and Claude.ai will generate a ready-to-save markdown file:

Personas worth trying:

• A patient mentor who asks before acting
• A no-BS startup advisor who challenges assumptions
• A punchy copywriter obsessed with clarity
• A kindergarten teacher explaining to a 5-year-old
• A strategist focused on jobs-to-be-done

Save the file, then quit and reopen Claude Code so it sees the new file. Activate it by typing /output-style and picking it from the list, or type /config and pick it under Output style. It kicks in from your next message.

Rather not save the file yourself? Paste the same prompt into Claude Code instead and add: "then save the file for me and tell me how to switch to it." Or use the walk-through prompt at the top of the page.

Or write it yourself

Five built-in styles

Switch between them inside Claude Code: type /output-style, or type /config and pick Output style:

• Default: standard Claude Code
• Proactive: Claude acts first and asks less; good when you trust it with routine calls
• Concise: shorter answers that lead with the result and skip the recap
• Explanatory: adds teaching insights to every response
• Learning: Claude leaves pieces of code for you to complete

Custom style file

Save as ~/.claude/output-styles/mentor.md:

---
name: Mentor
description: Explains decisions and asks questions before acting
---

You are a patient coding mentor. Before making changes:

1. Explain what you're about to do and why
2. Ask if there's a detail I should consider
3. After the change, point out one thing I could learn

Never write more than one file without pausing to check in.

Activate via /config → Output style. Tip: add keep-coding-instructions: true to the frontmatter if you still want Claude coding normally underneath the new voice.

📖 Official docs: Output styles ↗

Just pick a sound from the list

• Glass.aiff
• Submarine.aiff
• Ping.aiff
• Pop.aiff
• Tink.aiff
• Purr.aiff

This one doesn't need AI; the fun is picking which sound. Preview any macOS sound in Terminal first:

Swap the filename with any of the sounds above. Then drop the snippet below into settings.json.

Where it goes

Add to ~/.claude/settings.json (merge with any existing hooks section):

{
  "hooks": {
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "afplay /System/Library/Sounds/Glass.aiff"
          }
        ]
      }
    ],
    "Notification": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "afplay /System/Library/Sounds/Submarine.aiff"
          }
        ]
      }
    ]
  }
}

Stop fires when Claude finishes a turn. Notification fires when Claude is waiting for you.

Rather not edit the settings file yourself? Tell Claude Code which sounds you picked and ask it to add them to your settings for you, showing you the change first. Or use the walk-through prompt at the top of the page.

Tip: Pick quiet sounds (Tink, Pop) for frequent events. Save dramatic ones (Submarine, Sosumi) for Notification only.

📖 Official docs: Hooks reference ↗

Just type /config theme=dark

This one's a one-liner, not a config puzzle. Inside Claude Code, run:

Swap dark for any theme in the table under Go further, or just type /config alone and pick from the menu. Changes instantly. No restart. (Typing /theme works too, and that's where you can design your own custom theme.)

Pro tip: use auto if you switch between dark at night and light by day. Use light if you take screenshots; dark themes photograph poorly.

Or set it permanently

All theme options

Or set it permanently in ~/.claude/settings.json:

{ "theme": "dark" }

📖 Official docs: Settings reference ↗

Ask Claude.ai to write the command file

Describe what you want and Claude.ai will generate a ready-to-save file:

Commands worth building:

• /slack-recap: summarize my last 3 Slack messages into one action
• /decide: pick between two options with a pros/cons frame
• /elevator: turn a long idea into a 30-second pitch
• /reframe: offer 3 alternatives to a negative thought
• /translate: turn business-speak into plain English

Rather not save the file yourself? Paste the same prompt into Claude Code instead and add: "then save it for me." Or use the walk-through prompt at the top of the page.

Or write it yourself

Three steps

1. Create the folder: mkdir -p ~/.claude/commands
2. Save a file named after the command, e.g. ~/.claude/commands/joke.md
3. Type /joke in Claude Code

Template

---
description: Tell a joke about a topic
argument-hint: "[topic]"
---

Tell me a funny, clever joke about $ARGUMENTS.
Smart, not corny. One joke only.

Then type /joke quantum computing and Claude runs the prompt with "quantum computing" as the argument.

Tip: Commands in ~/.claude/commands/ work across every project. Commands in ./.claude/commands/ (inside a project folder) only work in that project.

2026 update: custom commands were merged into Claude's skills system. Your command files keep working exactly as shown, and when a shortcut outgrows one prompt, it can graduate into a full skill. That's tip P3.

📖 Official docs: Slash commands ↗