Claude CodeCSV export with empty rows, bad dates and mismatched column names
A SKILL.md the agent loads whenever you hand it a raw export — same cleanup, every time, no re-instructing.
Track 6 · Advanced
6.3
Skills, tools & extensions
Teach your AI new tricks and plug in tools
1Overview
The extensibility layer every AI agent shares — learn it once, carry it across tools.
The extensibility layer that any agent shares — skills, slash commands, subagents, MCP tools and hooks. What each one is, when to reach for which, where to find ready-made ones, and how to write your own SKILL.md so it works across tools. → Unlike the per-tool courses (which teach this inside Claude Code or Hermes), this chapter is the tool-agnostic hub: learn the ideas once and carry them everywhere.
1.1After this chapter you can
→Tell skills, commands, subagents, MCP tools and hooks apart — and pick the right one
→Find ready-made skills and MCP tools in the community catalogs and registries
→Read and write a SKILL.md from scratch — and know why it works across tools
→Turn a one-off solution into a reusable skill (Hermes-style) and share it
1.2When to reach for it
When you want to customise or extend an agent: add a capability, automate a routine, or connect it to your own systems.
1.3Key parts
Five parts — hooks, subagents, MCP tools, slash commands, and skills — the same five across tools, so what you pick up on one AI agent carries to the next.
2Lessons 5
2.1 Extend an agent with skills, commands, subagents, tools and hooks
This lesson explains the five extension types that appear in every AI agent – skill, slash command, subagent, MCP tool and hook – and how they differ.
Identify the five extension types and their purposes
Try
Create a new skill named DataSummary. In its folder add a SKILL.md file containing:
Name: DataSummary
Description: Summarize CSV datasets into key statistics.
Instructions: Load the CSV, compute count, mean, median for each numeric column, and output a markdown table of results.Paste this text in the Skills & Tools panel, click Add New Skill, then confirm the folder creation. Verify that the skill appears in the list and that the SKILL.md file is recognized before moving on.
- Create a skill folder and add a SKILL.md file describing its name, description and instructions
- Define a slash command by saving a prompt under a short name such as /review
- Launch a subagent to run work in an isolated context and receive only its summary
- Configure an MCP tool to connect the agent to an external system like GitHub or a calendar
- Write a hook script that fires automatically on a chosen event, for example after each file edit
- You'll see Five names that used to blur together now snap into distinct jobs – so when a tutorial says “add a skill” or “connect an MCP server”, you know exactly what kind of thing it means
- Takeaway An agent is a core plus an extensibility layer; the five pieces — skill, command, subagent, MCP tool, hook — recur in every tool
- Check Of the five extension types, which one never requires you to invoke it directly because it fires automatically on an event?
2.2 Choose the right extension
Choosing between skill, command, subagent, MCP tool and hook, or deciding to add nothing at all.
Do this first Extend an agent with skills, commands, subagents, tools and hooks
Pick the appropriate extension for a given need and know when to add nothing
Try
I keep forgetting to cite sources in my papers; I want the AI to automatically format citations according to our style.Paste the sentence into the Add Extension input on the skills-and-tools screen and hit Classify Need. Verify that the tool suggests a skill, not a command or MCP tool.
- Identify the nature of the need and decide which category it belongs to – skill, command, subagent, MCP tool or hook
- If the need is reusable knowledge, create a skill; if it is a repeatable prompt you type yourself, add a command
- For work that should run separately from your main thread, enable a subagent; for data or actions in an external system, configure an MCP tool
- When the action must fire automatically on a specific event, set up a hook, otherwise default to restraint and use no extension
- You'll see You can instantly identify whether a new annoyance requires a skill, a command, a subagent, an MCP tool, a hook, or no extension yet
- Takeaway Match reusable knowledge to a skill, shortcuts to commands, isolated work to subagents, system access to MCP tools and event‑driven actions to hooks – when unsure, start with plain prompts
- Check Which extension fits a reusable piece of knowledge, and which fits a prompt you type yourself?
- Cost The cheapest extension is the one you don't add. Each one you do add is paid for in context on every session, so the decision guide is also a budget guide.
2.3 Find ready‑made skills and tools
Three public catalogs host ready‑made AI skills and MCP tools: the anthropics/skills GitHub repo, agentskills.io and the MCP registry.
Do this first Extend an agent with skills, commands, subagents, tools and hooks
Locate the main catalogs that host reusable AI skills, MCP servers and community tool indexes
Try
List all skill names from agentskills.io that are categorized under Design, and output them as a bullet list.Paste the prompt into the Input field of the skills‑and‑tools interface and hit Run. Watch for any pagination notice – you may need to click Load more to see the full list.
- Open GitHub and navigate to the anthropics/skills repository to explore example skills and the template
- Visit browser and go to agentskills.io to read the SKILL.md specification and view the showcase of compatible tools
- Launch browser and open the MCP registry at registry.modelcontextprotocol.io to search for available servers
- You'll see A short bookmarks list – anthropics/skills, agentskills.io, the MCP registry – that covers most “is there already a skill/tool for this?” moments before you build anything
- Takeaway You rarely start from scratch; browse official repos and registries first, vet sources, then only create missing skills or tools
- Check Where do you look before building your own, and what do you read in a skill's repo before trusting it?
2.4 Create a reusable skill
A skill is a folder containing a SKILL.md file with required frontmatter (name and description) and Markdown instructions.
Do this first Extend an agent with skills, commands, subagents, tools and hooks·Find ready‑made skills and tools
Produce a SKILL.md that works across all Agent‑Skills tools
Try
Create a folder named tidy-dataset and inside it add a file called SKILL.md with the following contents:
---
name: tidy-dataset
description: use this when you need to clean a raw CSV export into analysis‑ready form
---
# Tidy dataset
- Drop any empty rows.
- Normalise all dates to ISO 8601 format (YYYY‑MM‑DD).
- Convert column headers to snake_case.
- Remove leading/trailing whitespace from each cell.
Save the file.In the skills-and-tools interface, click New Skill, then paste the above text into the editor that appears and hit Save. Ensure the frontmatter block is bounded by three dashes on its own lines; otherwise the skill won’t be recognised.
- Create a new folder for your skill inside the default skills directory (e.g., .agents/skills/
) - Add a file named SKILL.md inside that folder
- Write frontmatter at the top of SKILL.md with a name and a concise description that starts with “use this when…”
- Enter the step‑by‑step instructions in Markdown below the frontmatter
- Optionally add supporting subfolders such as scripts/, references/ or assets/ for extra resources
- You'll see The SKILL.md appears in the tool’s skill list and is automatically loaded when you describe a matching task
- Takeaway A folder with a correctly formatted SKILL.md (name, description, instructions) can be written once and reused everywhere
- Check What in the frontmatter tells the agent when to load the skill, and where do the instructions go?
- Cost Skills are free files, but each one still loads its name + description into every session. Keep them few, one-job, and well-described so the agent picks the right one.
2.5 Create and test a reusable Agent Skill
An Agent Skill is a folder with a required SKILL.md file that describes metadata and step‑by‑step instructions for an AI agent.
You will be able to author a minimal skill, place it where a supported tool can discover it, and verify that the tool lists the new skill during discovery.
- Open a terminal and create a new folder named
my-skill. - Inside
my-skill, create a file calledSKILL.mdwith the following content: ``--- name: hello-world description: Responds with a friendly greeting. --- ## Instructions When asked to say hello, reply with "Hello, I am your new skill!".`` - Create an optional
scriptssub‑folder (e.g.,my-skill/scripts) and place any executable you might want later – for this lesson it can stay empty. - Copy the entire
my-skillfolder into a location that a supported agent scans, such as.github/skills/in a repository used by GitHub Copilot or Claude Code (as listed on page 2). - Open the AI tool (e.g., launch Copilot in VS Code) and ask it to list available skills or trigger a discovery query like “What skills do you have?”
- You'll see The agent responds with a list that includes
hello-worldand its description, confirming that the skill was discovered and is ready for activation. - Takeaway Skills are portable, version‑controlled folders; once placed in a recognized directory they become instantly discoverable by any Agent Skills‑compatible tool without additional registration.
3You’ll know it worked 106 checkable outcomes in this chapter
- ✓Agent removes empty rows, normalises dates, and standardises column names in the exported CSV
- ✓The assistant flags ledger lines over 10% variance and drafts explanation requests
- ✓An approver runs "/approve-invoice" and receives a structured approval note that can be pasted into ERP
- ✓A proposal draft appears in the output with executive summary, problem statement, solution fit, and pricing for Acme Corp
- ✓A closed ticket that doesn't match an article appears in the FAQ gap log
- ✓The workflow completes quickly and produces expected test outputs
- ✓The regex matches the string HERMES.MD and counts it as an overage trigger
- ✓A new commit appears in the git log with the supplied message and includes all intended changes
106 outcomes in all — one per recipe below.
4FAQ, Tips & How-to 203
one problem, one solution, one action
Research & data tools5
Need up‑to‑date headcount by department with quarterly growth alerts
An assistant that answers "what is current headcount by department and how has it changed this quarter?" from the live HRIS export.
Need current candidate pipeline info
An assistant that answers "where are we on the senior designer search?" from live ATS data, not a week-old export.
Can't read every closed‑deal note
A win/loss pattern report — top 3 win themes, top 3 loss reasons, most-mentioned competitors — produced from your deal notes without reading each one yourself.
Need the right battle‑card info for a competitor
A skill that returns the right battlecard content for a named competitor, including the objection responses most relevant to the stated deal context.
Knowledge & docs8
I have to manually copy citations
An assistant that answers from your own library and pulls real references — not invented ones.
Support tickets need documented answers
An assistant that answers support tickets by citing the actual article, not its training-data guess.
Employees asking HR policy questions
Policy answers grounded in your own documents, with the section cited, so employees and managers trust the reply.
Redlined contract shows odd clauses
A clause-by-clause redline review that flags deviations from your standard terms, with a reference to the closest signed precedent.
Need a reply for an angry customer asking about a delayed refund
A skill that picks, personalises, and returns the right macro for any ticket in one step — faster than searching the macro list yourself.
Raw developer changelog
A customer-ready release note produced by one command, in a consistent format customers recognise, every release.
Closed tickets that lack matching FAQs
An automatic FAQ gap log that grows every time you resolve a ticket, turning repeat resolutions into a backlog of articles to write.
Need drug interaction or ICD‑10 code while charting
A quick, cited lookup during a note or chart review instead of switching to a separate reference app — the agent surfaces the data, the clinician makes the call.
Content & marketing8
Got monthly metrics but no time to write an update
A skill that turns "here are this month's numbers" into a ready-to-edit update in your established format.
Need every customer reply to sound like our brand
A skill that makes every drafted message match your brand voice without you re-describing it.
Captions lose brand look
A reusable brand kit the agent applies to captions, briefs and posts without being reminded each time.
Only have a video title?
A skill that emits a consistent, ready-to-design thumbnail brief from just the video title.
Want a ready‑to‑post job description from a few role details
A skill that turns "senior backend engineer, 5 yrs exp, Berlin, hybrid" into a ready-to-post JD matching your template.
Raw forecast numbers
A skill that turns raw forecast deltas into board-ready variance commentary in your established style.
Need a custom first‑day intro for a new hire
A complete day-1 welcome bundle — Slack message, checklist, agenda — produced in one command, personalised to the role and team.
Need a proposal draft from a brief
A skill that produces a structurally complete, on-brand proposal draft from a brief — ready to personalise, not to write from scratch.
CRM & sales5
Want to know which deals went quiet this week?
An assistant that answers "which deals went quiet this week?" from live CRM data, not a stale export.
Deal stalls after demo
A skill that turns "deal at proposal stage, last touch 8 days ago, pain point: slow reporting" into a ready-to-send follow-up.
Having to type a deal note after each call
Every session ends with an auto-generated deal note appended to your log — ready to bulk-import or review.
Messy post‑call voice notes
A skill that turns a rep's post-call voice notes or bullet points into a complete, structured debrief ready to paste into the CRM.
Need an account rundown before a call
A structured meeting-prep brief assembled in one command, so a rep walks in knowing the account, the open actions, and the goal.
Internal tools & ops16
Forget to run a formatter after saving a file
Consistent output with zero reminders: the check happens on the event, not on your memory.
Team needs one shared code‑review checklist
A version-controlled `/review` the whole team shares, instead of each person's private prompt.
Long, noisy test runs flood the chat
A clean main conversation: the subagent absorbs the noise and reports just what failed.
Slack missing GitHub PR updates
An assistant that reads real GitHub activity and drafts the Slack update for it, grounded in both.
Need to finish every end‑of‑month task in order
A one-word trigger ("run month-close") that walks the full checklist in the right order, every time.
Need a consistent vendor‑approval check
A consistent vendor-approval gate the whole team can trigger in one word, with no checklist drift.
New hire onboarding runs in a subagent
An onboarding run that completes all steps in a subagent and surfaces only the ones that need human follow-up.
Customer emails are a mess
A skill that turns an unstructured customer email into a labelled, routed, SLA-stamped ticket summary.
Ticket thread is messy
A one-word command that turns a ticket thread into a structured escalation note engineering can act on immediately.
Too many audit files to sort
A completed audit evidence pack assembled in a subagent, with a gap report of items it could not locate.
Can't keep track of tax filing dates
Automatic deadline awareness on every doc the agent touches — no deadline slips past because you forgot to look.
Need a ready‑to‑review job offer
A ready-to-review offer letter generated by one command, with every variable filled and every required clause present.
Run the employee off‑board checklist and get only the steps that need my sign‑off
A complete offboarding run that surfaces only the steps requiring human sign-off, not a wall of task-by-task output.
Bullet notes on an employee’s work
A skill that turns a manager's bullet notes into a properly structured, evidence-framed performance review draft.
Messy support tickets
A one-command tool that turns a messy customer ticket into a properly structured bug report engineering will not bounce back.
Need a personalized reply for each payment‑outage ticket
A complete set of personalised incident reply drafts — one per ticket — returned by the subagent, ready to bulk-review and send.
Dashboards & analytics4
Need a repeatable Monday sales summary
A repeatable Monday summary that reads live data and lays it out the way you like.
Can’t see which ledger lines exceed the budget
An assistant that reads the live ledger and surfaces only the lines that need your attention.
Want current AP/AR aging without pivot tables
An assistant that reads live AP/AR data and produces an aging summary with the accounts that need action today.
Want your quota tracker to stay current without typing
An always-current quota tracker updated automatically each time a deal note is written, with zero manual entry.
Commerce & payments3
Need to auto‑approve or flag expense reports
A skill that runs your expense policy on any submitted report and returns a clear approve/flag/query decision with the exact rule cited.
Need a uniform way to approve invoices
A consistent invoice-approval record produced by a single command, ready to paste into your ERP or audit trail.
Need a refund decision that follows policy
Consistent refund decisions grounded in policy, not agent memory — with the exact rule cited so the customer can see the reasoning.
Forms, surveys & feedback3
Interview notes are messy
A skill that turns unstructured interview notes into a filled-in scorecard with a rating and evidence quote for each competency.
Survey responses keep coming
A continuously updated sentiment pulse that flags risk signals in real time, not just at quarterly review.
Raw CSAT survey comments
A CSAT theme report — top praise themes, top complaint themes, products mentioned, detractor quotes flagged — produced from any batch of survey comments.
Booking & scheduling1
Finding a time that works for all interviewers
A draft interview schedule and invite text generated from real calendar availability, ready to send in one confirmation.
Trackers1
Ticket export never flags looming SLA breaches
An automatic SLA watchlist appended to your ticket export every time it updates, with the at-risk tickets sorted by time remaining.
Customer & client portals1
Need to see a customer's ticket history while replying
A reply drafting assistant that knows the customer's history, open tickets, and account tier before it writes a single word.
▸How-to Everyone
Avoid overloading the LLM with irrelevant rules by limiting the root rule file to core project info
▸How-to Everyone
Working inside a folder and need only its rules
Provide relevant rules only when working inside a specific directory to keep the LLM focused
▸How-to Everyone
Can’t tell the assistant where to work in a huge repo
Help the agent understand which part of a complex layout to target by adding a brief map of subdirectories
▸How-to Everyone
Need the bot to know my latest code changes and docs
Give Claude up-to-date information about recent commits, unstaged changes, or Confluence docs before a session
▸How-to Everyone
Want rules to update automatically after each AI chat
Keep rules evolving by running a headless review that proposes changes to claw.md files after each session
▸How-to Everyone
Need prompt templates that load only when you edit certain files
Keep sessions lean by loading skill templates only when editing relevant files
▸How-to Everyone
Need to find the right symbol in a huge codebase
Enable precise definition and reference lookups in large codebases without relying on simple grep
▸How-to Everyone
Need a ready‑to‑run demo setup
Quickly replicate the demo setup on any repository with a single plugin installation
▸How-to Everyone
Need to launch a website without fiddling with FTP
You can launch the generated website without manual FTP or server setup
▸How-to Everyone
Need a custom lead‑capture form on your site
You can extend the live site with a custom backend form using plain English
▸How-to Everyone
Want image or video creation inside your code
Allow Claude Code to call Higgsfield tools directly
▸How-to Everyone
Want to insert an avatar into a page
Insert the chosen avatar into the layout using Claude Code CLI commands
▸How-to Everyone
Need a Discord bot to trigger Claude skills via DM
A Discord bot can be used to trigger local Claude skills via direct messages
▸How-to Everyone
Keeping my local notes up‑to‑date after each feature
Updating the local Claude.md after each feature reduces tokens for future work
▸How-to Everyone
No consistent look across generated UI components
Ensure all generated components follow a consistent visual style
▸How-to Everyone
Big project feels tangled
Divide a large project into parallelizable phases to speed development
▸How-to Everyone
My git commits get cluttered
Keep version control tidy and provide clear progress markers
▸How-to Everyone
Want to launch multiple wave sessions simultaneously
Execute multiple sub-agents simultaneously to accelerate implementation
▸How-to Everyone
Project cluttered with leftover template files
Eliminate unwanted template features automatically during skill installation
▸How-to Everyone
Need UI design assets
Produce a unique, consistent UI design quickly
▸How-to Everyone
Want to launch multiple agents fast without permission dialogs
Start parallel sessions without permission confirmation delays
▸How-to Everyone
Can’t keep wave plans organized
Keep phased plans organized and accessible for agents
▸How-to Everyone
Want a single window to control your agents
Launch the central interface for managing Claude Code agents
▸How-to Everyone
Want an agent that runs silently
Create an agent that runs without opening a new terminal window
▸How-to Everyone
Can't switch between manual approval and auto‑accept for agent edits
Switch between interactive approval and auto-accept for agent edits
▸How-to Everyone
Need to launch an AI agent that makes changes automatically
Launch an agent that makes changes without asking for confirmation
▸How-to Everyone
Want code to improve itself on a schedule
Create a cron-like job that automatically improves code every X minutes
▸How-to Everyone
Need a way to let big jobs run on their own until they’re done
Execute complex, multi-step projects unattended until finished
▸How-to Everyone
Running agents won’t stop
Quickly halt every agent and move them to the Completed tab
▸How-to Everyone
Agent stopped and you need to pick up where it left off
Continue work on a previously halted agent without losing state
▸How-to Everyone
Want a separate dev server for every UI mockup
Create isolated worktree branches for each UI design and compare them side-by-side
▸How-to Everyone
Need to link your Clay account for API access
How to link your Clay account to Claude Code for API access
▸How-to Everyone
Want 50 HVAC leads with enrichment and email copy
How to request 50 HVAC leads with enrichment and personalized email content
▸How-to Everyone
I need a multi‑step automation instead of a single answer
You can launch a Claude Code dynamic workflow by prefixing your prompt with Ultra Code
▸How-to Everyone
Need to run several AI helpers at once
Explicitly asking for a workflow tells Claude Code to create parallel subagents
▸How-to Everyone
Running a tiny batch first lets you confirm the workflow behaves correctly
▸How-to Everyone
Lots of repetitive tasks need handling
Dynamic workflows can launch hundreds of agents to process items concurrently
▸How-to Everyone
For single-step tasks, a normal session is cheaper than a workflow
▸How-to Everyone
Multiple agents edit the same repo
When many agents edit the same repo, instruct them to work on separate branches
▸How-to Everyone
Want to keep a workflow for later use
You can persist and reuse a workflow by saving it through the UI
▸How-to Everyone
Need a step‑by‑step OWASP code audit
A practical workflow can scan code for each OWASP category and output markdown reports
▸How-to Everyone
API calls are hitting Anthropic
Learn how to redirect Claude Code's API calls from Anthropic to Minimax with a single JSON change
▸How-to Everyone
I want a Reddit clone project plan
Execute a full project plan generation with Minimax in under two minutes by using Claude Code's planning mode
▸How-to Everyone
Text gets lost over moving hero video
Use an inward-masking gradient so overlay text stays legible over video
▸How-to Everyone
Mobile layout is broken
Let Claude Code tweak responsive design for mobile devices automatically
▸How-to Everyone
Unsure which AI model and quality to use for code edits
Select the language model and quality before issuing edits
▸How-to Everyone
Need to adjust objects in a game level with voice or text prompts
Use Claude to add or adjust objects in the level
▸How-to Everyone
Discuss mechanics and aesthetics without changing the level
▸How-to Everyone
Need stronger AI reasoning
Select the Fable 5 model to unlock higher reasoning power in Claude Code
▸How-to Everyone
Complex tasks need more thinking
Setting reasoning effort to "extra high" balances quality and token usage
▸How-to Everyone
Using Opus for review agents reduces token cost without hurting quality
▸How-to Everyone
Need a tidy place for thumbnail assets
Create a project folder with subfolders for images and logos in Claude Code
▸How-to Everyone
Need OAuth tokens for Gmail and Slack
Learn how to authenticate external services for use inside a routine
▸How-to Everyone
I have an n8n JSON workflow and need a Claude routine
Convert existing no-code workflows into Claude routines with minimal effort
▸How-to Everyone
Need linked notes that stay in sync for AI memory
Organize notes into concept sections so the AI can follow links and auto-update memory
▸How-to Everyone
Need Claude to edit my project files
Marking a folder as trusted lets Claude create, modify, and run files inside it
▸How-to Everyone
Prefixing a filename with @ tells Claude to load the entire file as context for the current query
▸How-to Everyone
Want Claude to remember a task across chats
Using /goal lets you define a long-running target that Claude pursues across multiple interactions
▸How-to Everyone
Need to run a prompt every few minutes
The /loop slash command automates recurring tasks by re-executing a prompt at set intervals
▸How-to Everyone
I need multiple AI helpers at once
Sub-agents let you launch parallel Claude workers, each handling its own task and producing separate outputs
▸How-to Everyone
Want a simple view of local files
Claude can build a small web-based dashboard from folder data and launch it locally for immediate viewing
▸How-to Everyone
Need brand-specific guidance for AI output
Populate the context folder with brand-specific markdown files to keep AI output authentic
▸How-to Everyone
Want reusable prompts and auto actions
Store prompt templates and define skills that combine context, templates, and Higsfield actions for automation
▸How-to Everyone
Want to add the newest GLM model without hassle
You can have GLM-5.2 running in Claude Code within minutes by supplying an OpenRouter API key
▸How-to Everyone
Need relational tables for Lerty
You can instantly create a full database schema with realistic data using Claude Code
▸How-to Everyone
Not sure which AI agent to sell
A free bundle accelerates niche selection and sales outreach
▸How-to Everyone
I need a way to manage my coaching clients
You can quickly set up a full coaching system with minimal effort
▸How-to Everyone
Knowing this lets you avoid accidental billing when using Claude Code
▸How-to Everyone
Understanding this helps you recognize why certain patterns cause charges
▸How-to Everyone
Need to start a sub‑agent from the UI
You can spawn a new sub-agent directly in Claude Code
▸How-to Everyone
Need a custom AI assistant defined in a markdown file
You can define a custom agent by writing a markdown file with YAML front matter
▸How-to Everyone
Need to call another agent from your prompt
You can fire a sub-agent by including its trigger phrases in the main prompt
▸How-to Everyone
Precise, unique descriptions reduce accidental agent activation
▸How-to Everyone
Need a sub‑agent to stay within one project
Choosing project memory keeps a sub-agent's context limited to that project
▸How-to Everyone
Can't tell which agent wrote a response
Assigning a color lets you quickly spot which agent produced a response
▸How-to Everyone
Running tasks in sub-agents prevents token pollution in the main chat
▸How-to Everyone
When multiple agents respond to the same prompt
You can force a specific sub-agent to run by naming it in the prompt
▸How-to Everyone
Want my custom AI agents in a folder that syncs with git
Custom agents live in a *.claud folder, making them portable and trackable
▸How-to Everyone
Unreliable sub‑agent triggers
Repeatedly refining the agent's description makes it fire only when intended
▸How-to Everyone
Need to bundle a system prompt, tool access, and custom skills
A harness gives the agent controlled environment and capabilities
▸How-to Everyone
Need a ready‑to‑run mobile app skeleton
Create a full mobile app skeleton with one natural-language command
▸How-to Everyone
Need to route AI code calls via a local proxy
Point the CLI to the running proxy so requests go through it
▸How-to Everyone
Fast mode breaks some models
Turn off Claude Code's fast mode to avoid API errors
▸How-to Everyone
Want diverse critique on an idea
Get a council of personas to stress-test ideas, yielding a verdict and cheapest 48-hour test
▸How-to Everyone
Want a YouTube analytics dashboard
Using a slash-goal prompt lets you ask the model to build complex artifacts like a YouTube analytics dashboard
▸How-to Everyone
Want a prompt to run by itself on a schedule
Use /loop to run a prompt automatically on a defined interval
▸How-to Everyone
Need an agent to keep working nonstop until a goal is met
Keep an agent working continuously until a specified goal is achieved
▸How-to Everyone
Want to control your workspace from your phone
You can start a remote session by typing /remote control in Claude Code to get a QR code and link
▸How-to Everyone
Need to end a remote session quickly
You can end a remote session from either device, but it may not always succeed reliably
▸How-to Everyone
Want to use the same coding session on your phone
Scanning the QR code opens the same Claude Code session on your phone
▸How-to Everyone
Need to accept code changes on your phone
You can approve or deny suggested changes directly on the phone without using a laptop keyboard
▸How-to Everyone
My laptop and phone stay on while I remote in
Using the flag keeps both interfaces running so you can switch between laptop and phone
▸How-to Everyone
Can’t type on the go
You can use your phone's voice input to send prompts to Claude Code while away from the desk
▸How-to Everyone
Switching between different project folders on my phone
You can start separate remote sessions in different project folders and switch between them on the phone
▸How-to Everyone
When I want a script that’s just a single line per function
Agents output minimal, runnable scripts instead of verbose boilerplate
▸How-to Everyone
Can't get Ponytail into my project
You can add ponytail to any Claude Code project by placing its files in a hidden .cloth directory
▸How-to Everyone
Need concise answers just once
You can enable or disable concise mode for a single query without restarting Claude Code
▸How-to Everyone
Skill won’t auto‑load on startup
If the skill isn't auto-loaded, you can force it by interrupting and asking for it
▸How-to Everyone
Responses are too long and waste tokens
Using the skill cuts response length, saving tokens and speeding up iterations
▸How-to Everyone
Need a command‑line style personal portfolio
Use a pre-written prompt to instruct Claude Code to build a CLI-style personal portfolio
▸How-to Everyone
Need typing animations and one‑time tool call visuals in terminal
Extend the prompt to include typing animations, response streaming, and tool-call visuals that run once per session
First 20 primes printed — observable result of Claude Code task
When the task succeeds you see the first 20 prime numbers displayed in your terminal
Complex task needs step‑by‑step guidance
Listing sub-tasks as a numbered sequence makes Claude execute them in order
Letting Claude explore first gives context, and stopping with Escape prevents wasted work
Describe change
Providing a clear description of the desired refactor prompts Claude to output a step-by-step implementation plan
You can leave plan mode at any time if you decide not to proceed with the proposed changes
Changed something you didn’t mean to
Rewinding lets you recover from a bad change without re-typing everything
Checking the context lets you monitor how much information Claude retains at any moment
Using Tab speeds up command entry and reduces typing errors
Arrow-up lets you quickly reuse recent slash commands without retyping
You can quickly see which files have uncommitted changes without running git yourself
Need to stage files and make a commit
You can let Claude handle staging and committing with a descriptive message in one turn
Need a correctly named feature branch without typing git commands
You can create a correctly-named branch without typing git commands yourself
Want to batch‑review AI edits without prompts
Shift+Tab puts Claude into a mode where it edits without prompting, letting you batch-review changes with `git diff`
Need to open a pull request from the terminal
Claude can open a pull request on your remote repository directly after committing
Choosing a model each new session
Setting a default model avoids having to switch each time you start a new session
Reduces token usage by keeping full tool specs out of context until Claude actually calls the tool
Can’t access design docs stored in Google Drive while coding
Enables Claude to read files stored in Google Drive directly during coding, without manual copy-paste
Big test runs flood the chat
Using subagents for big test suites or log processing prevents large outputs from cluttering the main chat
Want to run a procedure instantly
Team members can run the full procedure with a single `/skill-name` call, eliminating copy-pasting
Manual steps get missed
Hooks turn optional manual steps into reliable, automatically-run commands, removing reliance on Claude's memory
Need a single, non‑interactive Claude reply
Run Claude Code non-interactively to get a single response and then exit
Seeing what files and messages Claude has loaded helps you understand its decisions before approving changes
You get an evidence-based decision about graduating by counting how many of six signals your app meets
You can defend your stay or graduate choice concisely by referencing the exact signal(s) that drove the decision
I want to tag each task with a low, medium or high priority badge
Describing a precise, single change lets the AI generate focused code modifications
Refreshing the app preview lets you see whether the AI-generated change works as expected
Locating the saved snapshot confirms the change was recorded correctly
Need proof a new feature works
Capturing a screenshot provides visual proof that the code change functions as intended
▸How-to Everyone
Tired of re‑typing the same prompt every time
A slash command lets you fire a prompt instantly by name, saving re-typing effort
▸How-to Everyone
When I need something to run automatically on a specific event
Hooks let the agent react automatically when a specific event occurs, without manual prompting
▸How-to Everyone
Delaying the addition of skills, subagents, MCP tools, or hooks saves context tokens and avoids unnecessary complexity
▸How-to Everyone
Need to check external skills and servers for safety
You reduce risk by reviewing source quality and limiting permissions before integration
What exactly is Claude Code, and is it different from the regular Claude chatbot?
Claude Code is a coding assistant that runs inside your computer's terminal (command line), not a browser chat window. Unlike the regular Claude chatbot, it can directly read your project files, edit them, run programs, and use tools like Git — all through plain English instructions. Think of it as an AI colleague who sits in your project folder and can do real work on your files, rather than just giving you text advice to copy-paste.
What is MCP, and do I need to worry about it as a beginner?
MCP (Model Context Protocol) is an open standard that lets Claude Code connect to external tools and data sources — for example, reading files from Google Drive, querying a database, or pulling data from other apps. As a beginner doing local data analysis you do not need MCP at all. It becomes useful later if you want Claude to interact with services beyond your own computer.
Does Claude Code work in VS Code or other editors, or do I have to use the terminal?
Claude Code is available in multiple environments: the terminal (full CLI), a desktop app for macOS and Windows, a VS Code extension, JetBrains IDE plugins, and a browser. For someone new to the terminal, the desktop app or VS Code extension may feel more comfortable. All environments use the same underlying engine, so your CLAUDE.md instructions and settings work everywhere.
How do I get the best results from Claude Code as a beginner?
The official docs recommend being specific: instead of 'fix this script', say 'the script crashes when the CSV has missing values in column 3 — fix it and re-run the analysis'. Start a session by asking 'what does this project do?' so Claude understands your files before making changes. For risky changes, use Plan Mode (Shift+Tab twice) to have Claude propose a plan you can review before any files are touched.
What are Claude Code's limitations for scientific work? Should I trust its output without checking?
No — always review Claude Code's output before using it in a paper or analysis. It does not make scientific methodological decisions for you and can miss subtle statistical errors or domain-specific edge cases. It works best on clearly defined tasks with verifiable outputs ('run these tests', 'clean these column names'). The more precisely you specify your question, the more reliable the result. Treat its output as a first draft written by a capable but non-specialist assistant.
How much does using Claude Code actually cost in practice?
On a Pro or Max subscription, Claude Code usage is included within your plan's usage limits rather than billed per token. 'Fast mode' (priority inference) consumes credits faster, so heavy or long sessions on the lower plans can hit limits. For occasional use — a few analysis sessions a week — the Pro plan is usually enough; monitor your usage and step up to Max only if you regularly run out.
Do I need a paid subscription to use Claude Code? Which plan should I get?
Yes — Claude Code requires a paid Claude plan; the free plan does not include access. The Pro plan ($20/month) is the entry point and works well for occasional use. If you hit usage limits regularly, the Max plan starts at $100/month (5× more usage) or $200/month (20× more usage). For occasional data analysis scripts, Pro is usually sufficient to start.
How do I install Claude Code? Do I need to be a programmer to do it?
Installation uses a single command you paste into your terminal, and the native installer is self-contained — you do not need Node.js or Python pre-installed. After installing, type 'claude' in your terminal and follow the browser login prompt. The official Quickstart walks through it step by step.
Can I use Claude Code for data analysis even if I don't know how to code?
Yes. You describe your dataset and your question in plain English, and Claude Code writes and runs the analysis for you. One published demonstration had it analyze decades of weather records and produce charts and written summaries with no manual coding by the user. The key skill you supply is understanding what question you are asking — not how to write the Python or R to answer it.
What kind of data analysis tasks can Claude Code help with?
Claude Code can load data files (CSV, Excel, etc.), clean messy data, run statistical analyses, generate plots, write Python or R scripts, debug errors, and produce written summaries of findings. It is well-suited to tasks in biology: analysing sequencing results, cleaning field-survey spreadsheets, running regressions, or automating repetitive file processing. It does not replace your scientific judgement about which statistical method is appropriate — that remains your responsibility.
Does Claude Code remember what my project is about between sessions?
By default each session starts fresh. However, two mechanisms carry knowledge across sessions: a file called CLAUDE.md that you (or Claude) write in your project folder with project-specific instructions, and auto memory, where Claude saves learnings like build commands and your preferences. Run /init when you start a project to have Claude generate a CLAUDE.md automatically from your files.
How is Claude Code different from ChatGPT for coding and data analysis?
The biggest practical difference is that Claude Code runs inside your project and directly reads, edits, and executes your files, whereas ChatGPT in a browser requires you to copy-paste code back and forth. Claude's large context window means it can hold an entire research project in view at once, which is useful for multi-file analyses. Independent comparisons find Claude strong at generating efficient Python and at long-document analysis.
▸How-to Everyone
Want a local AI notebook but hate command‑line installs
Claude Code provides an interactive terminal that automates the installation of Docker, pulls the Open NotebookLM repository and configures it for you. This removes manual command‑line steps and lets you get a running notebook in minutes.
▸How-to Everyone
Want to double‑check AI‑generated code before it runs
Using plan mode (Shift + Tab in Claude Code or similar) makes the AI output a step‑by‑step execution plan before it writes code. Reviewing this plan catches logical errors early and prevents unwanted changes, especially for beginners.
▸How-to Everyone
Want a bias‑free pull‑request check
Running a review in a new AI session (cleared context) prevents the writer’s bias from influencing the reviewer, similar to an independent code audit. The dedicated /review_pr command automates diff extraction and issue comparison.
The same set on /recipes, filtered by tool and role.
5Videos 3
Extending AI Agents: A live demo of the GitHub MCP Server
A concrete, real-world look at the extend-your-assistant idea, against a tool every developer already knows.
The Model Context Protocol (MCP)
The canonical intro to MCP, straight from the people who created the standard the chapter is built around.
The Missing Link: How MCP Servers Supercharge Your AI Coding Assistant
The most direct match for this chapter: MCP as the extensibility layer that unlocks tools and external context inside a coding assistant.
6FAQ 4
Where can I find ready-made skills to download?
Start with Anthropic's open skills repo (github.com/anthropics/skills), which collects example skills across document, design, development and enterprise tasks, plus the format spec and a starter template. The agentskills.io site hosts the standard, a quickstart, and a showcase of tools that read the same format. Many tools and frameworks also ship their own skill galleries, and community "awesome" lists collect more. Whatever you download is someone else's instructions, so skim the SKILL.md before trusting it.
Where do I find tools (MCP servers) to connect?
MCP — the Model Context Protocol — is the open standard for connecting an agent to external systems, and there's an official, searchable MCP Registry at registry.modelcontextprotocol.io. Anthropic also maintains a repository of reference MCP servers (github.com/modelcontextprotocol/servers) — filesystem, fetch, git, memory and more — that are good first connections to try. Because an MCP tool can read your data and take actions, connect it read-only first and grant the narrowest access that works.
Does a skill I write work in more than one tool?
Yes — that's the point of the format. Agent Skills is an open standard (a folder with a SKILL.md containing a name, a description and instructions) that began at Anthropic and is now read by dozens of agents, including Claude Code, Claude, Cursor, Codex, OpenCode, Goose, GitHub Copilot and Gemini CLI. Write the SKILL.md once and a compatible tool can load it unchanged. The wiring differs slightly per tool (where skills live, how they're enabled), so check your agent's docs — but the skill file itself is portable.
How do I make my own skill — like Hermes does?
A skill is just a folder with a SKILL.md: YAML frontmatter giving a name and a one-line description of when to use it, then your instructions in Markdown (and optional scripts or reference files). The description matters most — it's what the agent reads to decide when to load the skill. Agents load skills by progressive disclosure: at startup they see only each skill's name and description, and they read the full instructions only when a task matches. A common pattern (Hermes makes this explicit) is to solve a task once, have the agent save the solution as a skill, review and approve it, then reuse it by name. For the hands-on, tool-specific walk-through, see the Hermes course ("Let it write its own skill") and the Claude Code course ("Package workflows as Skills").
7Glossary 12 terms
Show the 12 terms
Skills, tools & extensions
skill- A reusable capability you give an agent: a folder with a SKILL.md the agent loads only when a task matches, so specialised know-how does not clutter every session.
Agent Skills- The open standard for the skill format (a folder + SKILL.md). It began at Anthropic and is read by many tools, so one skill works across them.
slash command- A saved prompt you invoke by name (e.g. /review) instead of retyping it — a shortcut, not loaded-on-demand like a skill.
subagent- A helper agent that runs in its own isolated context, so noisy or parallel work does not crowd your main conversation; it returns just a summary.
MCP- Model Context Protocol — the open standard for connecting an agent to external systems (GitHub, a database, your calendar) so it can read real data and act.
MCP server- A small program that exposes one external system over MCP; you connect it to your agent to give it that tool.
hook- A script that fires automatically on an event (e.g. after every file edit) — for things that should happen without you asking.
progressive disclosure- How agents keep many skills cheap: at startup they read only each skill name and description, and load the full instructions only when a task needs them.
SKILL.md- The one required file in a skill: YAML frontmatter (a name and a description) followed by Markdown instructions the agent follows.
frontmatter- The small block of metadata at the top of a SKILL.md (between --- lines) giving the skill its name and description.
description- The one-line summary in a skill's frontmatter that tells the agent WHEN to use the skill — the single most important line to get right.
.claude/commands/- The folder where Claude Code reads project slash commands — one Markdown file per command, checked into git so the team shares them.
8See also
💬 Discuss this chapter
Ask, share, or report — over on the Heidelberg AI community forum.