Claude Code Workflow Guide

Claude Code Workflow Guide ========================== This guide is for people who want to use Claude's newest capabilities—browser automation, file operations, repeatable workflows—but aren't developers. You don't need to know how to code. You do need to be comfortable following instructions and copy-pasting commands. If you've been using Claude through the chat interface and wondered why people keep talking about "Claude Code," this is your on-ramp. Claude Code isn't about writing software. It's about giving Claude the ability to actually do things—click buttons in your browser, organize files on your computer, run multi-step workflows—rather than just talk about them. The name is misleading. The capability is real. This guide is organized so you can skip what you don't need. If you only want browser automation (filling forms, extracting data from websites, managing tabs), you can stop after Part 2 and never touch the terminal. If you want the full capability set—file operations, compounding workflows, Chrome integration—Parts 1 and 3 will get you there. The workflows in Part 7 show concrete examples of what's possible once you're set up. Start with the "Pick your path" section to figure out which pieces you actually need. Then come back to the specific workflows when you have a task in mind. This isn't meant to be read cover-to-cover—it's a reference you return to. --- If you can copy and paste text, you can use Claude Code. This guide shows the safest, simplest way to set it up and start using it—without needing to be a programmer. --- What you're actually setting up ------------------------------- You'll use Claude in three possible "surfaces": 1\. Claude in the Claude app Best for: writing, summarizing, meeting recaps, document analysis. This is what most people already use. No setup needed beyond your account. 2\. Claude in Chrome (the side panel extension) Best for: doing tasks in your browser—clicking, filling forms, pulling info from websites. Claude in Chrome is now available in beta for paid Claude plans (it previously ran in limited preview). It doesn't require any terminal or coding knowledge. 3\. Claude Code (the terminal tool) Best for: working with files and folders on your computer—renaming PDFs, organizing docs, generating reports, running multi-step workflows. This requires using Terminal, but you'll mostly copy and paste commands. You can use only #2 and get significant value. Claude Code (#3) is optional. Pick the path that matches your comfort level. --- Before you start: two safety rules ---------------------------------- Rule 1: Start supervised. When Claude is controlling a browser, choose "Ask before acting" until you're comfortable. This means Claude will show you what it plans to do and wait for your approval before doing it. Rule 2: Don't test on important accounts. Avoid banking, payroll, admin consoles, and anything irreversible until you've practiced on low-stakes tasks. Start with something where mistakes don't matter. --- Quick start: pick your path --------------------------- Path A: "I don't want to use Terminal" Do this: - Install Claude in Chrome (Part 2 below) - Use the browser workflows in this guide - You're done. Skip Part 1 entirely. Path B: "I'm okay copy/pasting a few commands" Do this: - Install Claude Code (Part 1 below) - Optionally connect it to Chrome (Part 3) - Run claude, log in, and try one workflow --- Part 1: Install Claude Code (only if you want terminal) ------------------------------------------------------- ### What is "Terminal"? Terminal is a simple app built into your computer that lets you run short text commands. Think of it like texting instructions to your computer instead of clicking around. You'll mostly copy and paste the commands from this guide. ### Which account do you need? Claude Code works with: - Claude Console (requires billing set up at console.anthropic.com) - Claude App Pro or Max subscription - Enterprise providers (Amazon Bedrock, Google Vertex AI, Microsoft Foundry / Azure AI Foundry) If you have a Pro or Max subscription, you're set. If you're on the free Claude tier, you'll need either (a) Pro/Max or (b) Claude Console billing/credits to use Claude Code. --- ### macOS (Apple computers) Step 1: Open Terminal - Press Command + Space - Type Terminal - Press Enter A window with a text prompt appears. This is Terminal. Step 2: Install Claude Code Copy this entire line, paste it into Terminal, and press Enter: curl -fsSL https://claude.ai/install.sh | bash What you should see: Text scrolling by as it downloads and installs. When it finishes (usually under a minute), you'll see your prompt again—a line ending with $ or % waiting for your next command. Step 3: Verify it worked Copy and paste these two commands, one at a time: claude doctor claude --version - claude doctor checks if your setup is healthy and shows any issues - claude --version confirms what version you installed If you see "claude: command not found": - Close Terminal completely - Reopen it - Try claude --version again Step 4: Start Claude Code You run Claude Code from inside whatever folder you want it to work on. For your first test, just run it from your home folder: claude This starts an interactive Claude session in your terminal. Step 5: Log in (first time only) Inside Claude Code, type: /login A browser window opens. Complete the login there. Once authenticated, you're ready to use Claude Code. --- ### Windows Windows has two main options. The simplest for most people is the native PowerShell installer. Option 1: Native Windows install (PowerShell) — recommended Step 1: Open PowerShell - Press the Start button - Type PowerShell - Click to open it Step 2: Install Claude Code Copy and paste this line, then press Enter: irm https://claude.ai/install.ps1 | iex What you should see: Text scrolling as it installs. When complete, you'll get your prompt back. Step 3: Verify it worked claude doctor claude --version Step 4: Install Git for Windows (required) Claude Code on Windows needs Git for Windows to work properly. 1. Download and install [Git for Windows](https://git-scm.com/download/win) 2. If Claude can't find Git Bash automatically, tell it where to look: $env:CLAUDE\_CODE\_GIT\_BASH\_PATH="C:\\Program Files\\Git\\bin\\bash.exe" Step 5: Start Claude Code and log in claude Then inside Claude: /login Option 2: WSL (Windows Subsystem for Linux) WSL runs a Linux-like environment inside Windows. It's often smoother for developer tools, but it's extra setup. If you're already comfortable with WSL, use the same curl install command from the macOS/Linux section. --- ### Linux The same installer works for Linux distributions: curl -fsSL https://claude.ai/install.sh | bash Then verify with claude doctor and claude --version. --- ### If you had an older Claude Code install Run this to migrate to the current native binary: claude install Also remove any old aliases or symlinks you may have created manually. --- Part 2: Claude in Chrome (most approachable surface) ---------------------------------------------------- ### What this is Claude lives in Chrome's side panel and can: - Read what's on a webpage - Click buttons and links - Fill out forms - Extract information into structured formats - Do all of this with your supervision This is the easiest way to start using Claude's agentic capabilities without touching Terminal. ### Setup (5 minutes) 1. Go to the Chrome Web Store and search for "Claude for Chrome" (or look for Anthropic's official extension) 2. Click "Add to Chrome" 3. Pin the extension so it's always visible (click the puzzle piece icon in Chrome, then the pin next to Claude) 4. Click the Claude icon to open the side panel 5. Sign in with your Claude account 6. Set it to "Ask before acting" — this is critical for safety ### Why "Ask before acting"? Websites can contain hidden instructions designed to trick AI agents. This is called prompt injection. When you choose "Ask before acting," Claude shows you what it plans to do and waits for your approval. Once you approve a plan, Claude stays within that plan—but still asks before any irreversible actions like submitting forms or making purchases. This prevents surprises. Anthropic also blocks certain high-risk site categories (banking, trading, etc.) by default—so if Claude won't work on a particular site, that may be why. You can switch to "Act without asking" later, once you're comfortable and working on trusted sites. But start supervised. ### What Claude in Chrome cannot do - It only works in Chrome (not Arc, Brave, Safari, Firefox, or other browsers) - It requires a paid Claude plan (Pro, Max, Team, or Enterprise) - It can use your existing browser login state. If you're not logged in (or hit a CAPTCHA), Claude will pause and ask you to take over. --- Part 3: Claude Code + Chrome integration (optional power mode) -------------------------------------------------------------- ### What this is This connects your terminal Claude Code session to your Chrome browser. Claude can: - Do "file work" in Terminal (renaming, organizing, processing) - Open Chrome and verify results in a real browser - Inspect actual page content, including console errors and DOM elements - Fill forms and check that they work correctly This is ideal for: - Web forms that are finicky or require verification - QA checks where you need to see real browser behavior - Any workflow where "did it actually work?" requires looking at a real page ### Requirements - Chrome browser - Claude in Chrome extension version 1.0.36 or higher - Claude Code CLI version 2.0.73 or higher - A paid Claude plan (Pro, Team, or Enterprise) ### Setup First, update Claude Code to make sure you have the latest version: claude update Then start Claude Code with Chrome integration enabled: claude --chrome Inside Claude, verify the connection: /chrome What you should see: A confirmation that Claude can communicate with Chrome. Note: You need to use \--chrome each time you start Claude Code, unless you enable it by default. Inside a Claude Code session, run /chrome and look for the option to enable Chrome integration by default. Limitation: Chrome integration does not work in WSL (Windows Subsystem for Linux). Use native Windows if you need this feature. --- Part 4: Skills (make workflows repeatable) ------------------------------------------ ### What's a Skill? A Skill is a saved set of instructions that Claude can reuse automatically. Think of it like: - A reusable checklist Claude follows - A template that triggers when relevant - A way to encode "how we do things" so it's consistent every time If you find yourself giving the same instructions repeatedly—meeting recaps, weekly reports, formatting rules, research templates—a Skill turns that into a one-time setup. ### How Skills work Unlike custom GPTs where you manually select which one to use, Skills activate automatically when Claude recognizes the context matches. You build the Skill once, and it applies whenever relevant. ### Enabling Skills in the Claude app 1. Go to Settings → Capabilities 2. Enable "Code execution and file creation" 3. Scroll to Skills and enable ### Creating a Skill (Claude app method) Custom skills are uploaded as a ZIP file containing a folder with your skill files. 1. Create a folder on your computer named after your skill (e.g., meeting-notes-skill) 2. Inside that folder, create a file called SKILL.md 3. ZIP the entire folder 4. Upload the ZIP in Settings → Capabilities → Skills ### Creating a Skill (Claude Code method, for teams) If you use Claude Code and want to share skills with your team via git: mkdir -p .claude/skills/meeting-notes Inside that folder, create SKILL.md: \--- name: meeting-notes description: "Process meeting transcripts into decisions, action items, and summaries" \--- \# Meeting Notes Processor \## Required inputs \- Meeting transcript (text file or pasted text) \- Attendee list (optional) \## Output format 1\. Decisions (with rationale) 2\. Action items (owner + due date) 3\. Risks identified 4\. Open questions 5\. Summary for distribution \## Rules \- Do not invent owners or dates not explicitly stated in the transcript \- Mark uncertain items as "uncertain" \- Include 3-5 supporting quotes from transcript for key decisions \## Proof step List transcript quotes that support each decision, so the human can verify accuracy. Commit this to your repo. Teammates who pull the repo get the skill automatically. ### The key to good Skills The description field is the trigger. Make it specific enough that Claude knows when to activate it. "Process meeting transcripts" is better than "help with meetings." Custom skills you upload are private to your account. Organization-wide sharing is available via Team/Enterprise provisioning. --- Part 5: Slack integration ------------------------- ### What this is You can tag @Claude in a Slack thread, and it creates a Claude Code session using that thread as context. Status updates post back to the thread. The thread becomes the prompt. ### Requirements - Pro/Max, or Team/Enterprise with Claude Code access (premium seats) - Claude Code on the web enabled for your account - Claude app installed in your Slack workspace - GitHub connected with at least one repo ### Setup 1. Admin installs the Claude app from the Slack App Directory 2. Users connect their Claude account 3. Configure Claude Code on the web and connect GitHub repos 4. Choose routing mode: - Code only: Claude routes all requests to Claude Code - Code + Chat: Claude routes coding requests to Claude Code, other requests to normal chat ### Current limitations - Works in channels only, not DMs - GitHub only for now (no GitLab, Bitbucket, etc.) - One PR per session - Claude Code routing is for coding tasks; for summaries and general PM work, use Chat (or enable Code + Chat routing) ### For non-coding Slack work If you want Claude to summarize threads, extract decisions, or do other non-coding work from Slack: - Use Claude in Slack directly (tag @Claude with your request) - Or copy the thread content into the Claude app --- Part 6: Claude Code on mobile ----------------------------- ### What this is You can initiate Claude Code tasks from your phone. The work runs in the cloud, and you review results later. This is delegation that travels with you—start something while you're away from your desk, review it when you're back. ### Current availability - iOS: Claude Code is available via the Claude iOS app in research preview - Android: Not yet documented for Claude Code—check release notes and product updates for current availability Tasks initiated from mobile run on Claude Code on the web. You're not running compute on your phone—you're delegating to a cloud session. --- Part 7: Workflows ----------------- For each workflow below, I'll cover: - When to use it (the situation that triggers this) - What you need (inputs and prerequisites) - Steps (what to do, with copy/paste prompts) - What done looks like (expected output) - Verification (how to check it worked correctly) --- ### Workflow: Form filling (supervised) When to use: You have repeated form entries—vendor intake, CRM updates, event registrations, procurement requests—and a spreadsheet or list of the data to enter. What you need: - Your data in a table format (Google Sheet, Excel, CSV) - The target form open in Chrome - Claude in Chrome set to "Ask before acting" Steps: 1. Open Claude in Chrome side panel 2. Open the form you want to fill 3. First prompt (planning only): Scan this form. List every field you'll need to fill and what data you'll map to it. Don't type anything yet—just show me the plan. 4. Paste one row of your data (or a screenshot of the row) 5. Claude proposes a mapping. Review it. 6. If the mapping looks right, tell Claude to proceed 7. Watch it fill fields. For submit/purchase/send buttons, Claude should pause and ask. 8. Before submitting, run this verification: Before submitting, summarize what you entered in each field as a checklist so I can verify. 9. Compare the checklist to the form visually. Then approve submit. What done looks like: A filled form that matches your source data, with a checklist summary you've verified. Verification: Never allow Claude to click "Submit" on the first run without the checklist verification step. After a few successful runs with the same form, you can streamline. --- ### Workflow: Form filling (batch mode with Claude Code + Chrome) When to use: Same as above, but you have many rows and want to process them faster, or the form is finicky and you want Claude to see actual browser errors. What you need: - Claude Code installed - Chrome integration enabled (claude --chrome) - A CSV file with your data - The form URL Steps: 1. Create a working folder and put your inputs.csv there: mkdir vendor-intake && cd vendor-intake 2. Start Claude Code with Chrome: claude --chrome 3. First prompt: Connect to Chrome. Open the intake form at \[URL\]. For the first CSV row, fill the form but stop before submit. Then print a field-by-field summary for my approval. 4. Approve actions as prompted 5. After the first row is correct: Repeat for remaining rows. After each row, pause before submit and wait for my "go." What done looks like: All rows processed, with a pause-and-verify step for each submission. Verification: Spot-check at least 2-3 submissions against source data after the batch completes. --- ### Workflow: Competitive research When to use: You need structured comparison of competitor pricing, features, or positioning across multiple websites. What you need: - List of competitor URLs or company names - A clear sense of what data points you want (pricing tiers, feature limits, positioning claims) Steps (Claude in Chrome): 1. Create a simple table structure for your output (columns you want) 2. Open competitor site #1 3. Prompt: Find the pricing page on this site. Extract: tier names, prices, and key limits for each tier. Return as a table and cite the exact URL you found this on. If any prices aren't listed explicitly, write "Not disclosed" rather than guessing. 4. Repeat for each competitor 5. Final synthesis: Combine these into one comparison table. Add a row noting which competitor has the best value for \[your specific use case\]. List 3-5 meaningful differentiators I should pay attention to. Steps (Claude Code + Chrome, faster for many sites): 1. Create competitors.txt with one URL per line 2. Start Claude Code with Chrome: claude --chrome 3. Prompt: Open each URL in competitors.txt in Chrome. Navigate to pricing, extract tier/pricing/limits into pricing.csv. Include the source URL for each row. If pricing isn't explicit, write "Not disclosed" and note what you checked. What done looks like: A CSV or table with consistent columns across all competitors, source URLs for verification, and explicit "Not disclosed" where data wasn't available. Verification: Manually check at least 2 competitor pricing pages against what Claude extracted. Websites change; agents misread UI. Trust but verify. --- ### Workflow: Meeting processing (post-meeting) When to use: You have a meeting transcript and need to extract decisions, action items, and create a summary for distribution. What you need: - Meeting transcript (from Zoom, Meet, Teams, Otter, etc.) - Clarity on who the audience is for the summary Steps: 1. Paste the transcript into Claude (app, Chrome, or Code) 2. Prompt: From this transcript, extract: 1\. Decisions made (with the rationale discussed) 2\. Action items (include owner and due date if stated; mark "unassigned" or "no date" if not) 3\. Risks or concerns raised 4\. Open questions that weren't resolved Then draft two summaries: (a) Internal recap for the team (can include candid notes) (b) External recap for clients/stakeholders (professional, focused on outcomes) Keep the tone direct. Don't pad. 3. Verification prompt: List 5 short quotes from the transcript that support the most important decisions, so I can verify you understood correctly. 4. Check the quotes against your memory of the meeting. If they match, the extraction is probably accurate. What done looks like: A structured breakdown (decisions, actions, risks, questions) plus two summaries ready to send, with quote-backed verification. Verification: The quote-check step is the verification. If Claude's supporting quotes don't match what you remember being decided, dig deeper before distributing. --- ### Workflow: Meeting prep (pre-meeting) When to use: You have a meeting coming up and want to walk in prepared with an agenda, attendee context, and sharp questions. What you need: - Meeting title/purpose - Attendee names and roles (if known) - Any relevant docs or prior notes Steps: 1. Prompt: I have a meeting: \[title/purpose\] Attendees: \[names and roles\] Create a one-page pre-read with: \- Context (what this meeting is about and why it matters) \- Desired outcomes (what we should walk away with) \- Proposed agenda with time allocations \- Key risks or tensions to be aware of \- 10 sharp questions I should be ready to ask or answer Format it so I can paste it into an email. 2. If you want attendee research and have LinkedIn open: Look at this LinkedIn profile. Summarize: their role, what they probably care about, and one question they're likely to raise in this meeting. What done looks like: A one-page document you can send as a pre-read or use as your own prep notes, with specific questions rather than generic ones. --- ### Workflow: File organization (invoice sorting) When to use: You have a folder of messy files (invoices, receipts, documents) that need renaming and organizing. What you need: - Claude Code installed - A folder with the messy files Steps: 1. Navigate to your folder: cd ~/Downloads/invoices claude 2. Dry run first (no changes): Look at all the files in this folder. Read the contents where possible. Propose a renaming and organization plan: \- New filename format: YYYY-MM-DD Vendor - Type - Amount.pdf \- Organize into subfolders by vendor DO NOT rename anything yet. Just show me the plan as a table. 3. Review the table. Look for: - Dates parsed correctly? - Vendor names extracted right? - Any files it couldn't read? 4. Fix edge cases in your instructions if needed 5. Execute: Apply the plan. After each rename/move, log what you did. Stop immediately if any file would be overwritten. What done looks like: Files renamed consistently, organized into folders, with a log of what changed. Verification: Spot-check a few files. Open them and confirm the filename matches the actual content (date, vendor, amount). --- ### Workflow: Expense categorization When to use: You have transaction exports (from credit cards, banks, expense tools) that need categorizing for budgeting or tax purposes. What you need: - Claude Code installed - Transaction export (CSV or similar) - Your category list Steps: 1. Put your transactions file in a folder and start Claude Code there 2. Prompt: Read transactions.csv. Categorize each transaction into one of these categories: \[list your categories\] Output a new CSV with columns: Date, Description, Amount, Category, Confidence (high/medium/low), Notes For any transaction you're uncertain about, mark confidence as "low" and add a note explaining why. Flag potential duplicates. 3. Review the output, especially "low confidence" items What done looks like: A categorized transaction list with confidence flags, ready for review or import into your accounting system. Verification: Sort by "low confidence" and review those manually. Check a random sample of "high confidence" items to make sure the categorization logic is sound. --- ### Workflow: Research synthesis (compounding notes) When to use: You have research notes scattered across files and want to synthesize them into something useful that builds over time. What you need: - Claude Code installed - A folder structure for your research Steps: 1. Set up a structure: research/ raw/ (drop new notes here) competitors/ customers/ insights.md (the synthesized output) 2. When you have new notes to process: cd research claude 3. Prompt: Look at any new files in raw/. For each one: 1\. Extract key insights 2\. Categorize (competitor intel, customer feedback, market trend, etc.) 3\. Move to the appropriate subfolder Then update insights.md: \- Add new insights under the right headings \- Deduplicate anything that overlaps with existing content \- Add a "What changed this week" section at the top Do not delete existing content—archive anything you're removing under an "Archive" heading at the bottom. What done looks like: An insights.md file that grows smarter over time, with clear history of what was added when. Verification: Read the "What changed this week" section. Does it match what you remember putting in the raw folder? --- ### Workflow: Decision log from Slack thread When to use: A decision happened in a Slack thread but needs to be documented properly. What you need: - The thread content (copy or export) Steps (in Claude app or Claude in Slack): 1. Copy the relevant messages from the thread 2. Prompt: Turn this Slack thread into a decision record: \- Decision: (what was decided) \- Date: (when) \- Participants: (who was involved) \- Rationale: (why this decision, what alternatives were considered) \- Risks or concerns raised: \- Open questions: \- Next steps and owners: Be conservative—if something isn't explicitly stated, mark it as "not discussed" or "unclear" rather than inferring. What done looks like: A clean decision record you can paste into your documentation system. Verification: Check that the "Decision" line actually matches what was decided, not just what was discussed. --- ### Workflow: Document analysis (contract review) When to use: You need to review a contract and extract key terms, risks, or specific clauses. What you need: - The contract (PDF or text) - Knowledge of what you're looking for Steps: 1. Upload or paste the contract 2. Prompt: Create a clause inventory for this contract. For each major section, note: \- Section number and title \- Summary of what it covers \- Anything unusual or non-standard Then specifically flag any concerning language in: \- Termination provisions \- Indemnification \- Limitation of liability \- Assignment rights \- Confidentiality \- Data security/privacy \- Governing law and dispute resolution Output: A table of clauses, plus a "Questions for counsel" section listing anything I should get a lawyer's opinion on. What done looks like: A structured overview of the contract with flagged areas for attention. Critical note: This is not legal advice. Use this for initial triage and to surface questions for actual lawyers. Don't sign anything based solely on Claude's analysis. --- Glossary -------- Terminal / CLI: A text-based app where you type or paste commands instead of clicking around. "CLI" stands for Command Line Interface—same thing. Command: A short text instruction you type or paste into Terminal. Folder / Directory: A container for files on your computer. Same concept as folders in Finder or File Explorer. Path: The "address" of a folder—where it lives on your computer. Example: /Users/yourname/Documents/invoices WSL: Windows Subsystem for Linux. A way to run Linux tools on a Windows computer. Git Bash: A tool that gives Windows a "bash" terminal, which is the standard type of terminal on Mac and Linux. Extension: A small add-on program for Chrome that adds features. Side panel: The panel that slides out from the right side of Chrome, where Claude in Chrome lives. Prompt: The instructions you give Claude. Also refers to the blinking cursor in Terminal waiting for your input. DOM: Document Object Model. The underlying structure of a webpage that Claude can inspect when using Chrome integration. --- Common problems and fixes ------------------------- "claude: command not found" Your computer can't find Claude Code yet. - Close Terminal completely - Reopen it - Try claude --version again - If still broken, run the install command again "Permission denied" (macOS) macOS sometimes blocks new tools. - Run claude doctor and follow any suggestions - You may need to allow the app in System Preferences → Security & Privacy Chrome integration not connecting - Make sure you started Claude Code with claude --chrome - Check that your Claude in Chrome extension is version 1.0.36 or higher - Run /chrome inside Claude Code to verify the connection - Chrome integration doesn't work in WSL—use native Windows instead Login isn't working - Make sure you have a Pro, Max, Team, or Enterprise Claude subscription, OR have billing set up at console.anthropic.com - Try /logout then /login again - Check that your browser isn't blocking the popup Claude is slow or timing out - Large files or complex tasks take time - Check your internet connection - For very large tasks, break them into smaller pieces --- What's next ----------- This guide covers the foundations. Once you're comfortable: - Build Skills for your most common workflows so you don't repeat instructions - Chain workflows—research feeds into documents, documents feed into summaries - Share Skills with your team via git so everyone works consistently - Explore the plugin marketplace in Claude Code for specialized capabilities—there's an official Anthropic marketplace installed by default; browse it with /plugin The goal isn't to automate everything. It's to automate the parts that are repetitive and predictable, so you can focus on the parts that require judgment.