Claude Code Desktop Playbook

I read your session history, settings and project folders on this PC. Here is what they show.

Other habits I found

• You run in Auto mode almost all the time. You used Plan mode 3 times, for the exam system and the desktop pet, and both plans came out detailed and clear.
• Claude used the built-in Browser pane 708 times to check its own pages. This is a strong habit. Keep it.
• Claude ran about 1,180 shell commands for you. Claude Code's checkpoints can't undo shell commands, so git matters more for you than for most people.
• You have no MCP connectors, plugins or personal skills yet.
• All 31 sessions ran on this PC. You haven't started a cloud session yet. This matters for Projects (chapter 06).
• While I wrote this, your 5-hour usage window was 47% used and your weekly window 29%.

Do them in this order. Each one makes the next easier.

1. Give each project a memory file

   All 16 film sessions started from zero, and Claude had to learn your pipeline again each time. A CLAUDE.md file in the folder is read at the start of every session, so Claude knows the pipeline before you type.

   How: chapter 05.

2. Match the model to the job

   On a Pro plan, running Opus for every small change uses your limits fast. Keep Opus for planning and hard bugs. Use Sonnet 5.5 for normal edits, and Haiku 5.5 for helpers.

   How: models and chapter 08.

3. Hand side jobs to subagents

   Research and code searches fill your main conversation with long output. A subagent does the job in its own space and returns only a short summary. Several can run at the same time.

   How: chapter 04.

4. Put every project in git

   Git gives you a real undo, and it unlocks /code-review, worktrees, cloud sessions and Projects. You already have git installed.

   How: your safety net.

5. Turn your film routine into a skill

   You repeat the same steps for each film. A skill holds those steps. It loads only when needed, and you start it with one slash command such as /explainer-film.

   How: skills.

Four choices before you press Enter

Each new session has four controls in the prompt area. You can change model and permission mode later, mid-session.

Permission modes

For youYour two Plan-mode sessions (the exam system and the desktop pet revision) gave you the clearest specs in your history. Use this pattern for every new app: Plan first, then Auto to build. Press Ctrl+Shift+M to open the mode menu.

Models and effort

Effort sets how hard the model thinks: low, medium, high, xhigh or max. On the 5.5 models the default is medium. Raise it for a hard bug and lower it for renames. Type ultrathink in one message to get a single extra-deep turn.

Press Ctrl+Shift+I for the model menu and Ctrl+Shift+E for the effort menu. The usage ring next to the model picker shows how full the conversation is and how much of your plan you've used.

For youYou like comparing models; your Grok and Qwen pelican pages show it. Do the same here: run one real task on Sonnet 5.5 and again on Opus 5.5, then decide whether you can see the difference.

Shortcuts worth learning (Windows)

Small habits that help

• Correct Claude while it works. Type the correction and press Enter. You don't need to stop it first.
• Ask side questions in a side chat (Ctrl+; or /btw). It can read the session, but nothing goes back into it.
• Click an HTML, video or image path in the chat. It opens in the Browser pane, which is a quick way to watch a rendered film.
• Comment on a diff. Click a line in the diff view, type, then press Ctrl+Enter to send all your comments.
• Start a new session for a new task (Ctrl+N). Type /compact when a session you need to keep gets long.
• Before you commit, type /code-review. It needs git.

Your safety net: git

Claude Code saves checkpoints of its own file edits, but it doesn't track files changed by shell commands such as deletes, moves, scripts or renders. Git does. The desktop app also needs git for worktrees, /code-review, cloud sessions and Projects.

Git 2.54 is already installed on this PC. Open a session in a project folder and paste this:

Turn this folder into a git repository. Add a .gitignore for node_modules, build output, and large media files (.mp4, .wav). Make a first commit. From now on, commit after each feature that works, with a short clear message.

The GitHub CLI (gh) isn't installed. The app uses it to show pull request and CI status. Install it when you start using GitHub: winget install --id GitHub.cli

A subagent is a helper that Claude starts for a side job. It gets its own fresh context window, does the job, and sends back a short summary. Your main conversation stays small and focused, and several helpers can work at the same time.

Built-in helpers

How to call one

1. Ask in plain words. This is the most reliable way: "Use 3 subagents in parallel to…"
2. Name a custom agent: "Use the fact-checker agent on script.md."
3. Type @ and pick the agent, if it shows in the list.

Claude also starts helpers by itself when a job fits an agent's description. In the desktop app you can watch them under ⋮ → Background tasks in the session title bar.

Use 3 subagents in parallel to research the history of [TOPIC]. Split by era: before 1900, 1900–1980, 1980–today. Each subagent returns 8 dated facts with a source link for each. Then compare their results, flag any facts that conflict, and write a 90-second narration script (about 220 words).

Use the Explore agent to map how the clipboard manager works in this project: which files, the main functions, and how data flows. Don't change anything yet. Report back in 15 lines or less.

Make your own subagent

A custom subagent is one Markdown file. Put it in C:\Users\lenovo\.claude\agents\ to use it in every project, or in <project>\.claude\agents\ for one project. This one checks the facts in your film scripts:

---
name: fact-checker
description: Checks dates, numbers and names in a script or article against web sources. Use after writing any film script or history text.
tools: WebSearch, WebFetch, Read
model: sonnet
---

You are a careful fact checker. For every claim that has a date, a number or a name:
1. Find at least two independent sources.
2. Mark it Correct, Wrong or Unclear.
3. For Wrong claims, give the correct value.

Return a table: Claim | Verdict | Fix | Source URL.

You can also just ask: "Create a subagent called fact-checker that…", and Claude writes the file. The first time you create the agents folder, start a new session so the app picks it up.

When not to use one

• Quick edits, and back-and-forth tweaking of a page's look.
• Work that needs everything you've said in the conversation so far.
• Research, code searches, fact checks, test runs and log reading: jobs that make a lot of output you don't need to see.

Bigger jobs: dynamic workflows

When a job needs many helpers, such as checking every file or reading many sources, say "use a workflow", or put the word ultracode in your prompt. Claude then writes a small script that runs the helpers. /deep-research <question> is a built-in example. On Pro, workflows stay off until you switch on Dynamic workflows in /config, and runs default to a small size (fewer than 5 agents).

Agent teams are a different feature, and they only work in the terminal CLI. The desktop app doesn't have them.

CLAUDE.md: rules every session reads

CLAUDE.md is a plain text file of facts and rules. Claude reads it at the start of every session in that folder.

• One project: <project>\CLAUDE.md
• All your projects: C:\Users\lenovo\.claude\CLAUDE.md
• Type /init in a session and Claude writes a starter file for that folder.
• Keep it short, under about 200 lines. Put facts and rules in it, not long procedures. Procedures belong in a skill.

For youStart with animation by ai, your busiest folder. Open a session there and paste:

Read this folder, the mg-styles-15 README, and two finished film folders. Then create a CLAUDE.md here that explains our film pipeline step by step, the files each step makes, where Python and ffmpeg live on this PC, the list of styles, the default length and size, and the rules I keep repeating. Keep it under 120 lines and show it to me before saving.

It should come out close to this. I drafted it from your film-studio mod, so check the steps against your real pipeline:

# Explainer films (animation by ai)

## What we make
- 90-second motion-design explainer films, 1080p, one topic per folder.
- Styles live in mg-styles-15/ (15 styles, e.g. 09-bauhaus, 10-synthwave, 20-pixel).

## Pipeline for every film
1. Research the topic and write the narration script.
2. Write src/cues.json (timed scenes).
3. Make the narration audio: out/audio.wav
4. Render the master .mp4, then a smaller <name>_web.mp4
5. Write UPLOAD.md (title, description, tags).
6. Record style, title and file names in film.json

## Tools on this PC
- Python: C:/Users/lenovo/AppData/Local/Python/bin/python.exe
- ffmpeg: the imageio_ffmpeg binary (path is in the film-studio mod)

## Rules
- Check every date and number with a second source.
- Never change another film's folder.

A personal file for all projects could look like this:

# About me
- I build explainer films, browser games and small desktop apps.
- I'm still learning. Explain your choices in simple English, in 3 lines or less.
- Windows 11. Use Windows paths.

# How I like to work
- For a new app or a big feature, show me a plan before editing.
- Test web pages in the Browser pane before saying "done".
- Commit to git after each feature that works.

Auto memory: notes Claude keeps

Say "remember that the exam client must run offline", and Claude saves a note to C:\Users\lenovo\.claude\projects\<folder>\memory\. Later sessions in that folder read it. To put a rule in CLAUDE.md instead, say "add this to CLAUDE.md".

Skills: a procedure Claude can run

A skill is a folder with a SKILL.md file of steps. Only a short description sits in Claude's context until the skill is needed, so skills cost almost nothing while unused.

You already have these on your account: pdf, docx, xlsx, pptx, docs, deep-research, skill-creator, setup-writing-style and the browser and computer-use skills.

To use a skill, describe the task ("make a PDF of this"), or type / in the prompt box (or + → Slash commands) and pick one.

Make your own: the easy way

Use the skill-creator skill to make a personal skill called explainer-film. When I run it with a topic and an optional style, it should: suggest 3 fitting styles from mg-styles-15 if I gave none, research the topic with 3 parallel subagents, show me the script before rendering, then follow the pipeline in the CLAUDE.md of my animation by ai folder and tell me each file it makes.

Or write the file yourself

---
name: explainer-film
description: Plan and produce a 90-second motion-design explainer film in the animation by ai folder. Use when the user asks for a new film, video or animation about a topic.
argument-hint: [topic] [style]
---

Make a new explainer film about $ARGUMENTS.

1. Read CLAUDE.md in C:/Users/lenovo/Desktop/AI/animation by ai for the pipeline and tools.
2. If no style is given, suggest 3 styles from mg-styles-15 that fit the topic. Ask me to pick one.
3. Use 3 subagents to research the topic by era. Each returns dated facts with sources.
4. Write the script (about 220 words for 90 seconds). Show it to me before rendering.
5. Follow the pipeline steps in order. After each step, name the file you made.

Then run it like this: /explainer-film history of the bicycle 09-bauhaus

Which one should I write?

There are four ways to keep work going when you're away from your desk.

Cloud sessions

1. Push the project to GitHub.
2. Start a new session. Set Environment → Cloud.
3. Pick the repository and branch, then describe the task.
4. Close the app if you like. Check back in the app, at claude.ai/code, or in the Claude phone app.

To move a local session to the cloud, click the caret next to the session title and choose Open in → Cloud.

• Cloud sessions can't see your local files, your .claude\skills folder or local MCP servers.
• They offer Accept edits, Plan and Auto. Bypass isn't available.
• Usage counts against your plan, with no extra charge for the cloud machine.

For youThe cloud suits cs-fps, super mario and the web parts of the exam system. It doesn't suit film renders, which need your local Python and ffmpeg, or Windows .exe builds.

Projects (beta)

A project is one long conversation with Claude. You send tasks as they come up. Claude starts a thread, usually a cloud session, for each task, runs them in parallel, and reports back. The Overview pane sorts threads into Ready for review, Waiting on you, Working, Landing, Idle and Resolved.

Your access status

• ✓Pro plan. Projects are in beta for Pro and Max.
• ✕Used a cloud session. Not yet. The rollout starts with accounts that have used cloud sessions.
• ✕Code on github.com with the Claude GitHub App. Not yet. Threads that work on code need it.
• ?"Projects" in your Code tab sidebar. Check after the steps below.

How to get it

1. Push one project to GitHub. The exam system is a good first choice because it's already a git repository.
2. Install the Claude GitHub App on that repository.
3. Run one cloud session on it, as described above.
4. Look for Projects in the Code tab sidebar. If it isn't there, the rollout hasn't reached you yet. Join the waitlist.

Create a project

In the sidebar, click Projects → New project and fill in:

• Name (the only required field)
• Goal, in one line, for example "Ship version 1 of the online exam system"
• Context: repositories, files, folders or Google Drive folders that threads should read

Then click Create project.

Set it up before your first batch (Pro-friendly)

1. Click the gear icon, then go to Project settings → Memory → Project instructions. Paste the template below.
2. In Project settings → General, set Thread model and Thread effort. New projects use Opus at high effort, which uses a Pro plan quickly. Try Sonnet 5.5 at medium.
3. Tell Claude: "Propose threads and wait for my go-ahead. Run at most two at a time."
4. Send one small, real task first. When it finishes, open the thread and check its work.

This project builds the online exam system (teachers upload exams as JSON, students take timed exams, objective questions are graded automatically).

- Branch from main and open one draft pull request per thread.
- Before you call work done, run the project's tests and paste the summary lines.
- If something you need is missing, such as a repository, a secret or an API, say exactly what is missing and stop. Don't guess or mock it.
- Commit and push work in progress often.
- Don't merge or change CI settings without asking me in the thread.

Run a thread on your own PC

You need this for jobs only your PC can do, such as building the Tauri .exe or rendering a film.

1. Go to Settings → Claude Code. Turn on Use this computer from your phone and claude.ai, and add the folder.
2. In the project conversation, choose + → Work locally and describe the task.
3. On the card that appears, click Allow once.

Cowork (in the cloud)

Cowork is Claude's mode for tasks that aren't code. On Pro, new Cowork tasks run in the cloud since Oct 6, 2026, and you can no longer choose to run them only on your computer. Tasks keep going when the laptop is closed, and they show up on the web and on your phone.

1. Open the Cowork tab. In newer versions, you may instead choose Cowork at the bottom-left of the message box.
2. Describe the task. To go back to normal chat, choose Chat.
3. To use local files, connect a folder. Claude can reach only the folders you connect, and only while the desktop app is open. It copies the files it needs to the cloud.
4. For a one-off task that shouldn't use or add to memory, turn off Memory in the + menu first.

Dispatch lets you send tasks from your phone (Pro or Max). When a task is coding work, it becomes a Code session with a Dispatch badge.

For youGood Cowork jobs: turn your math modeling notes into a one-page cheat sheet, check a novel draft for timeline mistakes, or write a research brief for your next film and hand it to a Code session.

Read every file in my math modeling folder. Make a one-page cheat sheet: each method's name, what problem it solves, when to choose it over the others, and one short example. Save it as a PDF in the same folder.

If a task must stay on your PC only, use the Code tab with a Local session.

Remote Control: steer from your phone

• One session: flip the session's Remote Control switch, or type /remote-control.
• Every new session: Settings → Claude Code → Connect new sessions to Remote Control.
• Your PC must stay awake. You already have Keep computer awake while Claude works turned on.

Routines: run something on a schedule

1. In the Code tab sidebar, click Routines → New routine → Local.
2. Fill in Name, Description and Instructions. Pick the mode, model and folder.
3. Pick a schedule: Manual, Hourly, Daily, Weekdays or Weekly.
4. Click Run now once. Choose "always allow" on each permission prompt so later runs don't stall.

Local routines run only while the app is open and the PC is awake. After a missed run, one catch-up run starts when the PC wakes. Remote routines run in the cloud even when your PC is off. You can also just ask: "Set up a routine that runs every Monday at 9 am and…"

For youYou said Claude Code updates are hard to follow. Let a routine follow them for you:

Set up a local routine for this folder that runs every Monday at 9 am. It should read https://code.claude.com/docs/en/whats-new, find what changed since last week, and write a short summary in simple English to updates/YYYY-MM-DD.md. Focus on the desktop app, subagents, skills, mods, Projects and Cowork. Use Sonnet 5.5.

Mods

A mod is JavaScript or TypeScript that runs inside Claude Code and changes how it looks and behaves. It can add a pane, a band above the prompt, a new /command, or a rule that checks a tool call before it runs.

For youYou've already built one. film-studio in .claude\dev-mods adds /film <topic>, which briefs Claude on the whole pipeline, and /film-studio, which opens a pane listing the styles you've used and each film's stages.

Sample mods worth trying

These samples are from Anthropic's claude-code-playground repository.

Install the blast-radius sample mod from github.com/anthropics/claude-code-playground (claude-code/mods) on this PC. First read its code and tell me in 5 lines what it can do on my machine. Wait for my OK before installing.

To make a new mod, describe it: "Make a mod that…". Claude has a built-in skill for writing mods. To turn a mod off, disable its plugin under + → Plugins → Manage plugins.

MCP servers and connectors

An MCP server gives Claude new tools for reaching an outside service. A connector is an MCP server that you set up with a few clicks.

Your status: none connected, apart from the built-in Browser and session tools.

• Add a connector: click + next to the prompt box, then Connectors. This works in Local sessions. Manage connectors in Settings → Connectors or under Customize in the sidebar.
• Add a custom server: ask Claude to add it to the project's .mcp.json file.
• Online tutorials use claude mcp add …. That command needs the terminal CLI, which isn't installed on this PC. Connectors and asking Claude both work without it.

Recommended for you, in order

1. GitHub connector, once you start pushing repositories. Cloud sessions and Projects use GitHub anyway.
2. Context7 (third-party) gives Claude up-to-date docs for fast-changing libraries such as Electron, React, Tauri and Three.js. Add it to a project like this:
   <project>\.mcp.jsonCopy

   {
     "mcpServers": {
       "context7": {
         "type": "http",
         "url": "https://mcp.context7.com/mcp"
       }
     }
   }

3. Google Drive, only if your research or notes live in Drive. Cowork and Projects can read Drive folders.

• Skip Playwright or Puppeteer servers. The built-in Browser pane already does this job, and you use it a lot.
• Claude in Chrome works in your real Chrome with your logins. Use it when Claude needs a site you're signed in to.
• Computer use lets Claude click through native apps, such as testing your desktop pet. It's a research preview. Turn it on in Settings → This computer → System → Enable computer use.

Fewer is better. Each server adds tool names to Claude's context. Turn off the ones you don't use.

Plugins

A plugin is a package that can hold skills, subagents, hooks, MCP servers and mods. In a Local session, click + → Plugins → Add plugin to browse the official marketplace. Use Manage plugins to turn plugins on or off.

Output styles are built in and need no plugin. When you're studying, as in CC Learning, set Settings → Claude Code → Output style to Learning or Explanatory. Switch back to Default when you build.

Pro gives you a 5-hour window and a weekly window. When I checked, they were 47% and 29% used. Check them any time with the usage ring next to the model picker.

Habits that save the most

• One task per session. Every message resends the whole conversation, so long sessions get expensive. Your 31 transcripts add up to about 342 MB, around 11 MB each, which shows your sessions run long.
• After a break of more than an hour, the cache has expired and your next message reprocesses everything. Start a new session, or use /compact first.
• Ask quick questions in a side chat, not in the main session.
• Be specific. "Fix the jump height in player.js" costs less than "improve the game".
• Start small with Projects, workflows and many subagents. They use your plan fastest.
• If you hit the limit mid-task, Claude Code can wait for the reset and then continue.

One small step a day, each on your real work. Tick a box when you're done. Your ticks are saved in this browser only.

• Day 1Write CLAUDE.md for animation by aiUse the prompt in chapter 05. Read the result and fix anything wrong.
• Day 2Make one film with research subagentsCreate the fact-checker agent. Run the film session on Sonnet 5.5.
• Day 3Put your three busiest folders in gitStart with desktop pet, cs-fps and animation by ai. Then try /code-review on your next change.
• Day 4Turn the film routine into a skillUse skill-creator, then run /explainer-film on a new topic.
• Day 5Try Cowork on a non-code folderConnect your math modeling notes and ask for a cheat sheet.
• Day 6Push the exam system to GitHub and run a cloud sessionThen look for Projects in the sidebar, or join the waitlist.
• Day 7Automate and protectSet up the weekly Claude Code news routine, and try the blast-radius sample mod.

0 of 7 done

Where things live on this PC

Glossary

Your numbers come from the session history, settings and folders on this PC, read on Oct 9, 2026. Feature details come from the official docs on the same day. Features change fast, so when something here doesn't match your app, trust the docs.

• Desktop application
• Create custom subagents
• Dynamic workflows
• Projects (beta)
• Claude Code in the cloud
• Claude Cowork on web, desktop and mobile
• CLAUDE.md and auto memory
• Skills
• Mods overview
• MCP
• Anthropic's plugin marketplaces
• Models and effort
• Manage costs
• Scheduled tasks in Desktop
• Checkpointing
• What's new