Skip to main content

6. Agents File Access

Note: The video covers material not in the guide below — please watch in full.

Action Step

Complete this before moving on.

Copy the File-System-Demo folder into your own repo first. Then follow along with the video and run the three core prompts: fill the template from the transcript, rewrite it in brand tone, and add a First 30 Days section. Use WisprFlow for at least one prompt instead of typing. Once the brief is done, paste it into a Google Doc using the Markdown paste workflow from the video. Then go further — get creative and add something extra to the brief on your own.

Comment in Slack

Post your answer in your onboarding channel.

What was your biggest takeaway(s) from this training?

Training Guide

You've seen what Claude Code can do. Before we demo it, two quick settings you need to know.

Model: Always use the most powerful one available — right now that's Opus. Models are the engine behind everything. Don't default to a cheaper option to save a few seconds. Use the power. (Model names change — check the reference guide under this video for what's current.)

Permission Mode: Press Shift + Tab to cycle through three modes. Ask Before Edits means Claude asks your approval for every action — safe but slow. Plan Mode means it writes out a plan and waits for you to approve before running it. Bypass means it goes automatically — reads, writes, searches, whatever it needs. For knowledge work — writing, research, organizing — use Bypass. Save the guardrails for when a wrong move could break something. One more thing: while Claude is working, you can already type your next request. It queues up. You don't have to wait — stack your instructions and let it run.

Queue up Messages:

Now let's actually do the thing.

This is the training where it clicks. Where you stop watching and start driving. Because the single biggest difference between a consumer app and an agent platform isn't the settings or the models — it's this: the AI can read, write, and edit files on your computer.

Not in a browser. Not in a temporary chat window. On your actual machine, in your actual folders, alongside your actual work.

(Let's start with what that looks like)

The Demo: Client Research Brief​

Here's the scenario. You just got assigned a new client. You've got a transcript from the discovery call, a blank template you need to fill in, and a deadline. This is real Day 1 work at LeanScale.

In the Academy-Practice-Assets/File-System-Demo folder, you'll find three files:

  • client-research-template.md — a research brief template with empty sections
  • sample-discovery-call-transcript.md — a transcript from a discovery call with a fictional client called NovaPay
  • leanscale-brand-tone.md — our writing guide for client-facing work

Open the Academy-Practice-Assets/File-System-Demo folder in your explorer so you can see all three files. Now watch what happens.

COPY THE FOLDER TO YOUR REPO

Prompt 1: Read the Files, Write the Brief​

Here's the first prompt. You can type it or use WisprFlow:

"Read the client research template and the NovaPay discovery call transcript in the Academy-Practice-Assets/File-System-Demo folder. Fill in the template using the information from the transcript. Save it as novapay-research-brief.md in my Academy-Practice-Assets/File-System-Demo folder."

Press enter and watch.

The AI opens the template. It opens the transcript — a full conversation between four people, tangents and all. It reads both. Then it creates a new file — a filled-in research brief — and saves it exactly where you told it to.

Look at your explorer. The file is there. Open it. The sections are filled in. Company overview, funding, GTM stack, competitive landscape — all extracted from a messy conversation and organized into the template format. It pulled out the details Derek and Marcus mentioned, the phasing Sarah recommended, even the pricing page issue that came up as a side comment.

You didn't copy and paste anything. You didn't listen back to the recording. You didn't manually comb through the transcript for the relevant details. You pointed the AI at two files and told it what to do.

(That's the superpower. But we're just getting started)

Prompt 2: Rewrite It in Brand Tone​

Now look at the brief. It's accurate, but it reads a little generic. It doesn't sound like LeanScale.

Here's your second prompt:

"Read the LeanScale brand tone guide in the Academy-Practice-Assets/File-System-Demo folder. Now rewrite the NovaPay research brief to match our brand voice — direct, operator-minded, no fluff."

Watch what happens this time. The AI doesn't create a new file. It edits the existing one. And here's the key moment — you'll see diffs.

Diffs are the highlighted changes that show you exactly what the AI modified. Green means added. Red means removed. You can see every single word it changed. The vague language gets replaced with direct, specific language. The consulting-speak gets cut.

This is fundamentally different from a consumer app. In ChatGPT, you'd get a whole new wall of text and have to manually figure out what changed. Here, you see the edits the way you'd see tracked changes in a Google Doc — except the AI made them.

(One more)

Prompt 3: Add a Section​

The brief is solid. But you want to add a recommendation for the first 30 days — what LeanScale should tackle first based on the notes.

"Based on the transcript and the research brief, add a 'First 30 Days' section to the NovaPay brief with three specific recommendations for where LeanScale should start."

The AI reads the brief it already wrote, references the original transcript, and adds a new section at the bottom. More diffs — this time all green, because it's purely additive. It didn't overwrite anything. It added to the document like a collaborator would.

Three prompts. One file. Each prompt built on the last — template, then tone, then new content. That's the agent workflow.

How to Talk to It: The Prompt Pattern​

You just saw three prompts. Here's the pattern behind all of them:

What you want + Context + Where to put it

That's it.

  • What you want: "Fill in the template," "Rewrite this in brand tone," "Add a section with recommendations"
  • Context: The files it needs to read — template, transcript, brand guide. You give it paths or just tell it the folder.
  • Where to put it: "Save it as..." or "Write it to..." — the output path

For bigger tasks, add one more line at the end:

"Make a comprehensive to-do list for yourself to complete this task."

This makes the AI plan its own work. It'll break down the task into steps, show you the list, and check things off as it goes. You can watch the progress in real time.

The Copy Path Trick​

When you need to point the AI at a specific file, here's the fastest way:

  1. Find the file in your VS Code explorer
  2. Right-click it
  3. Click Copy Path
  4. Go to your prompt, press Shift + Enter to add a new line
  5. Paste the path

That's how you attach context. You can paste multiple paths — one per line — to give it several files at once.

Shift + Enter is your best friend here. It adds a new line in the prompt without sending the message. Build up your prompt, add your paths, then hit Enter when you're ready.

Images as Context​

It's not just files and text. You can paste images and screenshots directly into the Claude Code chat in VS Code, and the AI can read them.

Got a screenshot of a dashboard? A photo of a whiteboard? A mockup from Figma? Paste it right into the prompt. The AI sees what you see — it can describe it, extract data from it, reference it while working on your files.

Same pattern as before: What you want + Context + Where to put it. The context just happens to be a screenshot instead of a file path.

(Now there's one more thing you need to know — how to get your work out of Markdown and into the tools your team uses)

Getting It Out: Markdown to Google Docs​

Every file Claude creates is a Markdown file — that's the .md format you've been seeing. It's a simple way to write formatted text: headers, bold, bullet points, tables. It all works inside VS Code with MarkSharp rendering it for you.

But your team lives in Google Docs. Clients expect Google Docs. So how do you bridge the gap?

Step 1: Enable Markdown paste in Google Docs

  • Open Google Docs
  • Go to Tools → Preferences
  • Check "Automatically detect Markdown"
  • Click OK

You only do this once.

Step 2: Copy and paste

  • Open your .md file in VS Code
  • Select the raw text (not the MarkSharp rendered view — the actual text with the # symbols and ** marks)
  • Copy it
  • Paste it into Google Docs

The formatting carries over. Headers become headers. Bold stays bold. Bullet points stay bullet points. Tables work too.

That's the workflow: create in Claude Code, paste into Google Docs when it's time to share. You get the speed of AI-assisted writing and the collaboration features your team already uses.

What You Just Did​

Three prompts. One document that went from blank template to filled-in brief to polished, brand-aligned deliverable with strategic recommendations. The AI read your files, wrote to your files, edited what it wrote, and added to it — all without you leaving your workspace.

In a consumer app, that's at minimum thirty minutes of copy-pasting, tab-switching, and manual formatting. Here, it was a conversation.

This is file system access. This is why it matters.

(Next up: now that you know what the AI can do, let's talk about what's happening under the hood — how the AI actually works, and why that matters for getting better results)