Notes: Giving the Brain Hands

Lecture 9.1

What Can Claude Code Do With Your Apps?

Once an app offers a supported connection, you stop copying between apps and describe the result you want. That has changed how I choose software.

Notion on the left, Canva on the right, and Claude Code in the middle with a connection to each, so nothing gets copied by hand.

Key ideas

01Supported connection: An app has to offer one before Claude can work inside it.
02Describe the result: I say what I want, and Claude retrieves the source, follows my method and creates the material in the destination app.
03Willow Desk: Our fictional desk-organizer business, with 12 practice comments on a Notion page, carries every example in this section.
04Choosing software: When two apps do the same job, the one that lets Claude work inside it gives me a reason to pick it.

Why it matters. When two apps do the same job, the one that lets Claude work inside it gives me a reason to pick it. I moved from two CapCut team licences to Palmier Pro, then to DaVinci Resolve once Blackmagic offered an official MCP connection.

I still inspect the timeline and the finished video myself, because I need to know the edit does what I meant.

The whole section in one picture: twelve practice comments on a Notion page go through Claude Code and our feedback Skill, and come out as designs in Canva.

The same applies to software you build with Claude Code, without waiting for a directory listing. You can ask Claude to:

research how to expose its useful operations through MCP
build the connection
test one real action

Start from the job you want done, then look for a connection that gives Claude the actions that job needs.

Lecture 9.2

What Does a Tool Add to Claude?

A Skill tells Claude how to work, but only a tool can reach outside the conversation. Knowing which is which lets you read any Claude Code conversation.

Claude supplies the page address, a tool reads the Notion page through our account, and the twelve comments come back into the conversation where the Skill can use them.

Key ideas

01Tool: An action Claude can call, such as reading a page or creating a design.
02Skill: A method for how to analyze, plan or design once Claude has the material.
03Prompt: What I want, in my own words.
04One connection, several tools: Finding a page, reading it and creating a new one can come from the same connection, and Claude picks the one that fits.

When I read a conversation back, I can tell three jobs apart:

The prompt says what I want.
The Skill supplies a method.
A tool performs one action and returns something Claude can use.

Why the split matters. If the Notion page address is all Claude has, the Skill has nothing to analyze until a tool has read that page.

In the presentation request, the Skill plans the audience, the evidence and the slides, and the Canva create-design tool does the building.

Read a conversation as three layers: my request on top, the Skill as the method, and the Notion read as the action.

The same split shows up when the report becomes a presentation. The Skill decides the audience and how to arrange the evidence, and a Canva tool creates the design in my account.

Lecture 9.3

How Does Claude Reach Another App?

You will meet the words API, MCP server and connector on every app's setup page. Each means something specific, and they fit together in one chain.

Canva and Notion each have their own door for programs: Canva takes design content, Notion takes a page to read.

Key ideas

01API: The way one program asks another app to do supported things and gets results back. Each app has its own.
02MCP and MCP server: MCP is a common format describing an app's actions, what each needs and what each returns. The MCP server is the program that offers them.
03Connector: How a ready-to-connect service appears in the desktop app, including sign-in.
04Plugin: An installable package that can hold Skills, an MCP server or other helpers.

When I ask for a Canva deck, the request travels in four steps:

Claude picks the create-design action and supplies the content.
The MCP server passes the request to Canva.
Canva builds the design in my authorized account.
The result comes back for me to see.

Canva still decides which Canva actions exist, and Notion decides which Notion actions exist. MCP only gives Claude one common way to find and call them.

Claude calls the create-design tool on an MCP server, which makes the request to Canva's API inside my authorized account.

A plugin is a package: I read its contents to see what it adds, such as a Skill for the method and a connection for actions in Canva.

Where a server runs depends on the provider:

Local: a program that Claude Code starts on my computer.
Remote: a service that runs with the provider.

I follow the provider's setup instructions to see which one I have. Inside a plugin, the Skill supplies the method and the connection supplies actions in Canva.

Lecture 9.4

How Does Claude Choose a Tool?

Choosing a connection is a matching job: list the actions your work needs, then check each connection for them.

I look in the directory first, then the provider's own documentation for an official MCP server, and a maintained third-party connection last, after checking its source and required access.

Key ideas

01Connector directory: The first place I look. I check the publisher, the supported actions and the Claude Code setup instructions.
02CLI: A program driven by typed commands. Claude Code runs it in the terminal, waits and inspects the result.
03MCP or CLI: I take the supported route that does the job and fits my setup. When an app offers both, its current instructions decide.
04Tool Search: Claude finds the tools a request needs and loads their definitions then, not all of them at the start.

For the feedback job I need three actions:

Read a Notion page.
Save a report there.
Create editable designs in Canva.

I check every connection for those. A provider can publish an official MCP server in its own documentation without appearing in the directory. For a third-party connection, I check its source and required access first.

The same program can be reached by a typed command in the terminal or by an MCP tool, and Claude gets the video file either way.

About 19,500 tokens of MCP tools are loaded in this session, while 1,440 more tools wait outside the conversation until Claude searches for them.

My context panel showed where the space goes in one session:

Window: one million tokens, with about 383,800 in the conversation.
Loaded MCP tools: about 19,500 tokens, or 2 percent.
Deferred tools: 1,440, not in the conversation at all.

Those numbers describe one session, and returned page contents and reports still take space when they arrive.

Make it yours

Freelancer or consultant

Your input
A client's project notes in an app like Notion

Ask Claude
Find the connection for that app and read one page

Check
The action list includes reading and saving, and the page really comes back
Small business owner

Your input
The booking or invoicing software I already run

Ask Claude
Find out how Claude can work with it

Check
One supported route that does my job and fits my setup
On a team at work

Your input
The tool my team keeps notes or tickets in

Ask Claude
Which route do its current instructions recommend for Claude Code?

Check
What the first request returned

Lecture 9.5

How Do You Give Claude Access to Notion?

When Claude cannot read a page, the fix is almost always in the connection or the account, not in your Skill. Four checks tell you where to look.

Claude receives the Notion page as Markdown, the same plain-text table as our earlier feedback file, with F09 still in place.

Key ideas

01Connectors: In the desktop app I browse connectors, search Notion, choose Connect to Claude, sign in and approve workspace access.
02Turn it on: A connected service also has to be switched on in this conversation's connector list.
03Name the page: I give the exact page address so Claude knows which page I mean. A name works when only one page matches.
04Ask for IDs: I ask for the page title and every ID back, so I can check what was retrieved.

When a request fails, I check four things:

Can this session use the connection?
Did the right account sign in and authorize?
Can that account open the page?
Did the tool call return what I asked for?

If the account can't open the page, the connector won't give Claude access.

Four checkpoints for a failed request: the session, the signed-in account, whether that account can open the page, and what the tool call returned.

If the first attempt fails. Mine said Claude could not finish connecting. Once I was signed in to Notion, trying again worked.

Name the exact page. I turned Notion on in the conversation and gave Claude the page address. It returned the page titled Willow Desk customer feedback with the 12 fictional records F01 through F12.

Make it yours

Freelancer or consultant

Your input
A client's feedback page in Notion, shared with my account

Ask Claude
Read this exact page and list its title and every ID

Check
The returned IDs match the page
Small business owner

Your input
The page where I keep customer comments or reviews

Ask Claude
Read it and list the title and every comment ID

Check
The count and IDs match what I see when I open the page myself
On a team at work

Your input
A shared page such as meeting notes or a request log

Ask Claude
Read it, and if that fails, run the four checks

Check
Whether my own signed-in account can open the page

Lecture 9.6

Can Your Skill Use Data From Other Apps?

Your saved Skill is not tied to where its input came from. Point it at a page in another app and it still checks every comment.

Same Skill, new source: the report still covers every comment, 12 of 12, and keeps the single tray-edge comment visible.

Key ideas

01Same method, new source: The connection supplies the comments and the Skill supplies the analysis.
02Coverage: 12 comments in, F01 through F12, and every theme shows its IDs and quotations.
03The rare comment: F09 says the tray edge catches a finger. That is one possible problem to investigate, not a confirmed defect.
04Where the report goes: The source page, a new page or a subpage. I chose a subpage so the comments and the analysis stay together.

After Claude ran the Skill on the Notion comments, I checked three things:

Coverage against the page Claude retrieved.
The theme table, comparing each group with the original comments.
Whether F09 stayed visible.

The report keeps F09 visible even though setup and packaging complaints appear more often. It proposes actions, and what Willow Desk investigates first is still my choice.

Notion is both where Claude reads the comments and where the report is saved, with the unchanged Skill as the method in between.

Where the report goes. I asked Claude to save it as a subpage beneath the feedback page, and its reply included a link.

Other sources. If your feedback lives in Google Sheets or Excel, a supported connection may retrieve a table the Skill can analyze. Keep the original comments and IDs in that source so you can check the report.

Make it yours

Freelancer or consultant

Your input
A client's survey responses in Notion, each with an ID, and my saved analysis Skill

Ask Claude
Use this page as input and save the report as a subpage under it

Check
The coverage count matches the page, and every theme shows IDs and quotations
Small business owner

Your input
My customer comments, kept with IDs, and the analysis Skill I built

Ask Claude
Run the Skill on the page you retrieved

Check
A rare but worrying comment is still visible, and I decide what to investigate first
On a team at work

Your input
Feedback or ticket notes in Notion, or a table from a supported spreadsheet connection

Ask Claude
Run our saved Skill and put the report beneath the source page

Check
Each group against the original comments, before the team acts

Lecture 9.7

How Can Claude Reuse a Tutorial?

You can let Claude inspect tutorials you trust and save the method as a Skill, as long as you check what it examined and what it produced.

The Skill says to help the team decide and leave the decisions open, yet the revised deck's last slide still reads Start with F09, so I check the slides against the report.

Key ideas

01Watch: A GitHub project that prepares video captions and selected frames, so Claude sees the explanation and the examples on screen.
02Marketplace: A catalogue of packages Claude Code can install. I add one from a repository address, sync it, then add the plugin.
03My Watch choices: Local, so the work stays on my computer. Efficient, a smaller selection of frames. None for extra transcription, because both tutorials have captions.
04Sources reference: The note inside the saved Skill that names both tutorials and says how much of each Claude examined.

To turn two tutorials into a Skill, I did four things:

Installed Watch, which prepares captions and selected frames.
Picked Phil Waknell's presentation talk and a Canva presentation tutorial.
Asked Watch to inspect them and Skill Creator to save the lessons as Canva Presentation Designer.
Chose Local, Efficient and None for extra transcription when Watch asked.

Check the result. Claude proposed an outline and asked about a brand kit, and I approved the outline and skipped the kit. It pointed out that the revised final slide says to start with the F09 check.

I made the method and used it in the same conversation, so I have not tested whether Claude would find the new Skill on its own.

That priority is my decision, so a generated design can still depart from the Skill, and I compare it with the report.

Make it yours

Freelancer or consultant

Your input
Two tutorials I trust, such as a proposal talk and a Canva tutorial

Ask Claude
Inspect them with Watch and save the method as a Skill

Check
The sources reference, to see how much of each video Claude examined
Small business owner

Your input
A tutorial on the kind of post or flyer I make

Ask Claude
Inspect it, save the method as a Skill, then use it on my material

Check
The finished design against my own notes, before I rely on the method
On a team at work

Your input
A recorded talk on presenting, a design tutorial and the report we need to discuss

Ask Claude
Make a presentation with the saved Skill

Check
Any recommendation the slides make, because choosing the priority stays with us

Lecture 9.8

How Can Claude Use a Saved Design Method?

You can ask for a deck and a brief by who will use them, and leave the page sizes and layouts to Claude.

Each proposed investigation sits beside the comments behind it, with an empty cell where my decision goes.

Key ideas

01Plan first: Claude considers what the team should know, believe, feel and do after seeing the feedback, then proposes the slides and a portrait brief.
02Revise at the plan: If the decisions are already made, I say so. If the brief goes out by email, I ask for it in that form.
03Two formats: A short decision deck for the discussion and a one-page action brief to refer to afterwards.
04Edit in Canva: Opens the design in Canva's editor, after Canva asks me to sign in.

Claude showed its plan before it built anything, and the plan is where I can steer:

If the decisions are already made, I say so.
If the brief goes out by email, I ask for it in that form.
Otherwise I approve the plan, as I did, and skip the brand kit.

From that plan Claude made a 5-slide deck and a one-page brief in the same conversation. The deck's final slide leaves the order of the proposed actions for me to choose.

Edit in Canva. For the brief I chose Edit in Canva, signed in when Canva asked, and selected its table and text. The proposed investigations sit beside the supporting feedback IDs and quotations, with space for my decision.

Before choosing how many slides or pages to make, think about who will use them and what you want to help that person do.

Make it yours

Freelancer or consultant

Your input
A client's findings in Notion and a Canva Skill I saved earlier

Ask Claude
A short decision deck and a one-page brief, with the plan shown first

Check
The brief opened in the Canva editor, where I make my own changes
Small business owner

Your input
Notes for a month of social posts, in Notion or a spreadsheet

Ask Claude
Retrieve them, apply my saved method and prepare the designs in Canva

Check
The designs, which I revise in the same conversation
On a team at work

Your input
A report we already discussed, plus the decisions made

Ask Claude
Tell Claude the decisions are settled, and ask for a deck to explain them and a brief to keep

Check
The deck and brief against the report, with open choices left to the team

Lecture 9.9

Connect an MCP That Is Not in the Directory

An app can have an official MCP server even when it is missing from the directory, and the provider's own documentation is where you find it.

Meta's documentation gives the server address, and the check I suggest is one returned result: the account names and IDs Claude can see.

Key ideas

01Official, not listed: When the directory has no listing, I check the provider's own documentation.
02Desktop custom connector: Connectors, Manage connectors, Add custom connector. I name it, paste the address, and it appears under Yours.
03Terminal setup: I ask Claude to add the server without logging in. It becomes a local MCP for the project, and I authenticate myself.
04/mcp: Claude Code's list of MCP servers, with whether each needs authentication and how many tools it has.

Check the docs first. For Meta Ads, the connectors and plugins showed no listing from Meta, so I asked Claude to find the official documentation. The Ads MCP Server page gave the address, mcp.facebook.com/ads.

That one address works two ways:

Desktop app: add a custom connector, sign in, and approve the Facebook reconnect step myself. It shows Connected with a checkmark.
Terminal: after removing that connector, I asked Claude to configure the server without logging in. It adds a local MCP for the project, and I authenticate with /mcp.

My suggested first request only reads. It lists the ad accounts, before anything is planned or changed.

The terminal route does not show under Yours in the desktop app. In /mcp the row turned green with 97 tools, and Claude had made no tool calls yet.

Make it yours

Freelancer or consultant

Your input
A client's advertising account and the provider's documentation page

Ask Claude
Connect the official route, sign in myself, and ask whether Claude can access the MCP

Check
Connected and a loaded tool list, before any client data is touched
Small business owner

Your input
My own business's advertising account

Ask Claude
List the account names and IDs you can access

Check
The list matches the accounts I can open myself
On a team at work

Your input
The server address from the provider's documentation

Ask Claude
Add it as a desktop connector, or have Claude add it from the terminal while I authenticate

Check
The row in /mcp, showing the connection and its tool count

Lecture 9.10

Workflows You Can Run With Connected Apps

Every workflow here has the same three parts, so you can describe your own the same way: starting material, connected actions and a result you can inspect.

The template to copy: write down your starting material, the connected actions and the useful result.

Key ideas

01Creative: Duncan Rogoff starts from one product photo, a brand brief and shared visual references, so the images and video concepts look like they belong together.
02Advertising: Felipe Vergara gives account details and a strategy first. Missing identifiers had to be supplied, and some targeting still needed manual adjustment in Ads Manager.
03Browser: Elie Steinbock has Claude move through a new onboarding flow in his app and look for problems.
04Slack and docs: Read a named thread, list decisions and open questions, draft tasks. Context7 retrieves library documentation for a build.

For your own ad account, I suggest a smaller first task. Have Claude:

Read recent campaign results.
Compare them with a stated goal.
Prepare a report on which campaigns deserve attention.

The account data supplies the facts, and your Skill explains how to judge them.

A smaller first task for an ad account: the account supplies the facts, my Skill supplies how to judge them, and the report flags the campaigns that deserve attention.

Resolve with Blackmagic's official MCP: Claude assembles lectures, transitions and music while I inspect the timeline.

My own case is video. We used Resolve to assemble this section's lectures, apply transitions and background music, and edit out mistakes.

Blackmagic's official MCP gives Claude supported actions in Resolve, so I direct from Claude Code and inspect the timeline afterwards. My editing time for a whole section went from 10 days to 21 hours.

Lecture 9.11

Which Workflow Will You Connect?

The pattern is complete, so the next step is yours: pick one job you already do and connect it.

Describe your job to Claude Code and ask it to research the route: the directory, the provider's documentation, or a maintained third-party MCP or CLI.

Key ideas

01Describe your job: Where the information starts and what you want at the end.
02Ask Claude to research: Ask for the supported actions, the required setup and one result you could verify.
03Start small: One connected action, opened in the destination app and compared with your request.
04Agent: Takes responsibility for one part of a larger job and returns its findings to the main conversation.

To choose your first workflow, I would:

Pick a job I already do. The source might be a spreadsheet, the discussion might be in Slack, or the result might belong in a video editor.
Ask Claude Code to research useful connections for that workflow.
Take the recommendation that helps, and leave the others for later.

When the job grows, it may include researching the source, analyzing the findings and preparing the final material. An AI agent can take responsibility for one of those parts and return its findings to the main conversation.

Start with one small connected action, then open the result in the destination app and compare it with what you asked for.

As a job grows, an agent can own one part, such as researching the source, and return its findings to the main conversation.

Directing a larger job then means deciding three things:

which part an agent owns
what information it needs
what result should come back

Quick reference

Connecting

Desktop connector (Notion)	Connectors, Browse connectors, search Notion, Connect to Claude, sign in, approve workspace access. Then turn it on in the conversation's connector list.
Retry after a failed attempt	My first Notion attempt said Claude could not finish connecting. After signing in to Notion, the second attempt worked.
Custom connector (not in the directory)	Connectors, Manage connectors, Add custom connector. Name it, enter the provider's server address, choose Sign in when needed, leave automatic registration selected.
Meta Ads server address	mcp.facebook.com/ads, from Meta's developer documentation page for the Ads MCP Server.
Plugin marketplace (Watch)	Browse plugins, add a marketplace from the repository address, sync it, then Add the plugin.
Terminal MCP setup	Ask Claude to add the server and not log in. It adds a local MCP for the project and says authentication is still needed.
Authenticate a terminal MCP	In Claude Code enter /mcp, select the entry, choose Authenticate and continue in the browser. A green row with a tool count means connected (Meta Ads showed 97).

Commands

/mcp	Lists MCP servers, flags any that need authentication and shows tool counts. Connections imported from claude.ai appear in a separate group.
/watch:watch	Use it to run Watch on tutorial links you paste in. It asks Local or Gemini, a detail level, and what to use for videos without captions.

Where things live

Desktop custom connector	Under Yours in the connector settings, shown as a Web Custom connector.
Terminal-added MCP	In the project's local MCP list, a different place from the desktop connector list. It does not show under Yours.
Notion report	A subpage beneath the original feedback page, with a link in Claude's reply.
Canva Presentation Designer Skill	Created in the project, with a sources reference naming the tutorials and how much of each Claude examined.
Claude Ads repository	On GitHub, with skill folders for Amazon Ads, Google Ads and Meta Ads.

Words used in this section

Tool	An action Claude can call, such as reading a page or creating a design.
API	The interface one program uses to request supported operations from another app.
MCP	Model Context Protocol, a common format describing an app's actions, the inputs each needs and the result it returns.
MCP server	The program that offers those actions through MCP. It can run locally or with the provider.
Connector	How a ready-to-connect service appears in the desktop interface, including account sign-in.
Plugin / marketplace	A plugin is an installable package that can hold Skills, an MCP server or other helpers. A marketplace is a catalogue of packages.
CLI	A program driven by typed commands, which Claude Code can run through its terminal tools.
Tool Search	Claude finds relevant tools and loads their definitions when needed, not all at the start.
Browser MCP / Browser Use / computer use	Browser MCP is one particular browser connection. Browser Use is a separate project with its own MCP routes. Computer use is the broader ability to click, type and read the screen.

Use X when / Y when

Skill / tool	Skill for a method of analyzing or designing. Tool for an outside action, such as reading a page or creating a design.
Directory / official docs / third-party	Directory first, checking publisher, supported actions and Claude Code setup. Provider documentation when an official server is not listed. A maintained third-party connection as another route, after checking its source and required access.
MCP / CLI	MCP when the connection offers the action as a tool. CLI when a typed command does the job, such as generating a file. When both exist, follow the app's current instructions.
Local / remote MCP server	Whichever the provider's setup instructions name.
Exact page address / search by name	Address when you mean one page. Name when it is unambiguous, adding the title or address if similar pages cause confusion.
Desktop connector / terminal setup	Desktop for a manual route you do yourself. Terminal when you want Claude to configure it while you authenticate.
Watch Local / Gemini	Local to prepare captions and frames on your computer. Gemini to send the video to Gemini through an API key.
Watch Efficient / more detailed	Efficient for a smaller selection of frames. More detailed when the visual examples need closer inspection.
Watch None / WhisperX	None when the videos have captions. WhisperX to build a transcript from audio when they do not.

When a request fails

1. Session	Can this session use the connection?
2. Account	Has the intended account signed in and authorized access?
3. Permission	Can that account open the named page or perform the action? Changing the Skill will not fix this.
4. Result	Did the tool call return what you asked for? Check counts and quotations against the source.

Numbers from the recordings

Context window shown	One million tokens, with about 383,800 in the active conversation.
Loaded MCP tools	About 19,500 tokens, or 2 percent. Server instructions are a separate row, about 6,000 tokens, 0.6 percent.
Deferred MCP inventory	About 1.1 million tokens across 1,440 tools, available to find and not loaded into the conversation.
Willow Desk feedback	12 fictional records, F01 through F12. F09 is the tray-edge concern.
Canva outputs	A 5-slide decision deck and a one-page action brief.
Resolve editing time	10 days down to 21 hours for an entire section.

More in this section

Files and links

Section notes/resources/library/claude-code/notes-giving-the-brain-hands
Complete resource pack/resources/library/claude-code/giving-the-brain-hands-complete-resource-pack
Connected-app practice guide/resources/library/claude-code/connected-apps-starting-guide
Meta connection practice guide/resources/library/claude-code/meta-connection-practice-guide
Knowledge checks/resources/library/claude-code/connected-apps-knowledge-checks
Workflow sources/resources/library/claude-code/connected-app-workflow-sources
Feedback analysis Skill/resources/library/claude-code/customer-feedback-analyzer-skill