Five Things I'd Build With Claude in a Weekend

In the video I counted down five builds and saved the most important one for last. This is the full breakdown of how I built each one for myself, plus the order I'd build them in if I were starting this Saturday.

Here's the twist: I'd build them backwards. The second brain goes first, because every other build reads from it. A writing-style file is useless if Claude doesn't know who it's writing for. A competitor report is noise if Claude doesn't know what you sell. So you lay the foundation, then everything you stack on top gets better for free.

None of this needs you to be a developer. It's a folder of plain text files and a handful of prompts you can copy from this page.

Before you start: one folder, one home

Every build here is a few markdown files (plain text with a .md ending) sitting in one folder. By Sunday night yours will look something like this:

my-brain/
  CLAUDE.md                  the front door: who you are, what to read
  context/
    me.md                    build 1: the context layer
    work.md
    team.md
    current-priorities.md
    goals.md
    writing-style.md         build 2: how you write
  knowledge/
    how-we-work.md           build 3: the team knowledge base
    decisions.md
    sops/
  competitors/
    competitors.md           build 4: competitive intelligence
    snapshots/
    reports/
  content/
    ideas.md                 build 5: the content pipeline
    scripts/

You need somewhere for Claude to read that folder. Pick one:

• Claude Code (what I use). Claude works directly inside the folder on your computer and reads a file called CLAUDE.md at the start of every session, so it knows you before you type a word. It runs in the terminal, the desktop app or VS Code, and it's included on the paid Claude plans. If you want all five builds, this is the one.
• Claude Cowork. The same idea inside the Claude desktop app on a paid plan, no terminal. You connect a folder and it can read and write the files in it. Instead of CLAUDE.md, it uses folder instructions, so paste your CLAUDE.md text in there. (Anthropic is folding Cowork into the main Claude app, so the menus may look slightly different on your account.)
• A Claude Project on claude.ai. No install at all, and it works on the free plan. You paste your CLAUDE.md into the project's instructions and upload your context files as project knowledge. Great for builds 1 to 3. Builds 4 and 5 work better with a real folder, because Claude needs to save new files each week.

Pick one and stick with it. The files are the asset, not the app, so you can always move them later.

The weekend plan

That's the order in the video, flipped. Each section below covers what it is, how mine works, and how to build yours.

Build 1: The context layer (your second brain)

What it is. A small set of files that tells Claude who you are, what your business does, who's on your team, what matters this month and what it should never do. Without it, every chat starts from zero and the answers come out generic. With it, Claude walks into every task already briefed.

How mine works. My CLAUDE.md is about 120 lines. It says who I am, what the business sells and to whom, and that everything should support one thing: a great customer experience. Then it points at five short files in context/: me, work, team, current priorities and goals. It also links my writing rules, a decision log, and a folder per client and project. Every session reads all of it before I ask for anything. When my focus shifts, I update current-priorities.md and the whole system shifts with it.

Build yours.

1. Make the folder. Call it whatever you like.
2. Open Claude in it and paste this:

I'm building a context layer so you know me before every task. Interview me one question at a time. Cover: my role and how I work, the business (what we sell, to whom, at what price), my team (who owns what), my priorities for the next 30 days, my goals for this quarter, and what you should always and never do. If an answer is vague, ask a follow-up before moving on. When we're done, write five files: context/me.md, context/work.md, context/team.md, context/current-priorities.md and context/goals.md. Keep each one under a page, and use tables where they help.

1. Answer properly. Talk to it like a new chief of staff on their first day. This is the part people rush, and it's the part everything else depends on.
2. Create CLAUDE.md in the top of the folder and paste this in, then fill the gaps:

# [Your name]'s assistant

You are my assistant. Before any task, read these:

- @context/me.md (who I am and how I work)
- @context/work.md (the business, what we sell, who we sell to)
- @context/team.md (who owns what)
- @context/current-priorities.md (what matters this month)
- @context/goals.md (this quarter's targets)
- @context/writing-style.md (how I write)

## Always
- Ask before doing anything that sends, posts or deletes.
- Show your reasoning when a conclusion isn't obvious.

## Never
- Guess a contact detail, a price or a date. Ask.
- [Add your own]

In Claude Code, the @ lines pull those files in automatically. In Cowork, paste it into the folder instructions. In a Claude Project, paste the text into the project instructions and upload the context files as project knowledge instead.

Done when you can start a fresh chat, ask "what should I focus on this week?", and get an answer that could only be about you.

Keep it alive. A stale brain is worse than none, because Claude will confidently act on last month's priorities. Update current-priorities.md whenever your focus changes and goals.md at the start of each quarter. Ten minutes, and it pays for itself immediately.

Keep it short. Anthropic's own guidance is to keep a CLAUDE.md under 200 lines. Put the detail in the context files and keep the front door lean.

One rule: don't put passwords, bank details or anything you couldn't stand to see leaked into these files. Context, not secrets.

Go deeper: my original setup walkthrough, The AI Second Brain.

Build 2: Your writing-style file

What it is. One file that teaches Claude how you actually write, so drafts sound like you and not like AI slop.

How mine works. It's 25 lines. It covers format (tables and short paragraphs), a never-use list (em dashes, emojis, "circle back", "touch base", "moving forward"), a growing list of AI-tell phrases I've caught in drafts, my tone for my team versus clients, and sentence rhythm, with a before-and-after example. It ends on one default line: "shorter, sharper, more concrete." Every draft Claude writes for me checks against it.

Here's a trimmed piece of the real file:

## Never use
- em dashes (use commas, colons or a new sentence)
- emojis
- corporate-speak ("circle back", "synergy", "moving forward")

## Sentence rhythm
- Sharp does not mean staccato. Vary sentence length.
- Contractions are good. They read warmer.

## Default
When in doubt: shorter, sharper, more concrete. Strip filler.

Build yours.

1. Give Claude your real writing. The quickest way on claude.ai is the Gmail connector (Customize > Connectors), which lets Claude search and read your mail, so you can ask it to pull emails you've sent. Otherwise, drop 30-50 sent emails or documents into a writing-samples/ folder (copy-paste into a text file is fine).
2. Paste this:

Read my last 50 sent emails (or everything in writing-samples/). Write context/writing-style.md with five sections: Tone (how I write to clients versus my team), Never use (words, phrases and habits I avoid, including generic AI phrases like "I hope this finds you well"), Sentence rhythm (length, contractions, how I open and close), Examples (three short before-and-after rewrites in my voice), and Default (one line for when you're unsure). Quote my real phrasing where you can. Be specific, not vague.

1. Read it properly and edit it. Cut anything that isn't you. Add the phrases you can't stand that it missed.
2. Make sure CLAUDE.md points at it (the template above already does).

Done when you ask for a reply to a real email and you'd send it with one or two edits.

The habit that makes it great. Every time a draft sounds off, don't just fix the draft. Add the offending phrase to the never-use list. Mine has grown every week, and that's why it works.

Go deeper: the claude.ai-only version, step by step, in Emails That Sound Like You.

Build 3: The team knowledge base

What it is. How your business actually runs, written down: your SOPs, how the team works, the decisions you've made and why. In the video I called this the bridge between "cool demo" and "built around stuff I already do." Claude uses it as the reference point for any decision it helps you make.

How mine works. Three pieces:

• SOPs in references/sops/, one file per process.
• A decision log, one line per meaningful decision, append-only. It's short, but it stops Claude (and me) relitigating things we've already settled.
• A folder per client and project, each with a README that says the status, dates and key context. There are dozens of them now, and they're the first place Claude looks when I mention a client by name.

Build yours.

1. Start with three SOPs, the three processes you're most sick of re-explaining. Record a voice note or a quick screen recording explaining one, get the transcript, and paste it in with this:

Here's a transcript of me explaining how we [onboard a new client]. Turn it into an SOP at knowledge/sops/[client-onboarding].md: the purpose, when it applies, who owns it, the steps in order, the tools used, and the mistakes people usually make. Don't guess. List anything I skipped or left vague as questions at the bottom.

1. Write knowledge/how-we-work.md. How you communicate (which tool for what), how work gets assigned, what "done" means, and your response-time norms. Ask Claude to interview you for it, the same way as Build 1.
2. Start a decision log at knowledge/decisions.md. One line each, in this format:

[2026-10-02] DECISION: Workshops only sold inside a bigger engagement. | REASONING: One-offs took the same prep and never led to repeat work. | CONTEXT: Reviewed after Q3.

1. Add two lines to your CLAUDE.md:

- Before recommending how to do something, check knowledge/sops/ for an existing process.
- When we make a decision in conversation, offer to add it to knowledge/decisions.md. Never edit past entries.

If you record your meetings, this gets much faster. Paste a transcript and ask Claude to pull out every decision and every process someone explained, then file them.

Done when a new hire (or Claude) could answer "how do we onboard a client?" from the folder alone, without asking you.

Build 4: Competitive intelligence

What it is. A weekly read on what your competitors are doing: new offers, pricing changes, website changes, new ads, and which of their posts are blowing up.

How mine works. My tracker pulls every reel five competitors posted in the last month, flags the ones doing at least three times that account's usual views, and has a second AI model judge why each one worked. On the first real run it went through 185 reels in under two minutes for 49 cents, and found 23 outliers. Those outliers feed straight into my content ideas (Build 5).

Build yours. Start with websites, because they're the easiest win and Claude can read them directly with web fetch.

1. Create competitors/competitors.md:

# Competitors

| Name | Website | Pages to watch | Instagram |
|---|---|---|---|
| Acme Co | acme.com | /pricing, /services, / | @acme |
| Northwind | northwind.com | /pricing, /case-studies | @northwind |

1. Run this once to take your first snapshot, then once a week:

Read competitors/competitors.md. For each competitor, fetch every page listed under "Pages to watch" and save a clean text copy to competitors/snapshots/[name]/[page]-[today's date].md. Compare each page with its most recent previous snapshot and list what changed: prices, offers, headlines, new or removed pages. Then write competitors/reports/[today's date].md with what changed, why it might matter for us (read context/work.md for what we sell), and one thing we could do about it. If a page won't load, say so. Don't guess.

The first run has nothing to compare against, so it just takes the snapshot. From week two you get a real "what changed" report.

1. Ads. Meta's Ad Library is public and shows the ads any page is running right now. It's a heavy page that Claude often can't read reliably, so the honest weekend version is this: once a week, screenshot your competitors' active ads, drop the screenshots in, and ask Claude to summarise the offers, hooks and angles, and what's new since last week's report.
2. Their content. Tracking which posts are blowing up takes a scraper and a bit more setup. My exact setup is written up in full, linked below.
3. Make it weekly. Put it in your calendar for Monday morning and paste the prompt, or have it run on a schedule. Claude Code and Cowork both support scheduled tasks on paid plans. One catch: a task that reads your local folder only runs while your computer is awake and the app is open, so pick a time you're usually at your desk.

Done when you get a Monday report you'd actually read, and it has told you one thing you didn't know.

Go deeper: my full Instagram setup, with the numbers from the real run, in The Competitor Content Tracker.

Build 5: The content pipeline

What it is. Going from "what's working right now" to a script to an edited video, with as little manual grind as possible. Putting yourself out there only has upside, and the reason most people don't is the hours it used to take.

How mine works. I'll be honest about this one, because the "one click" version took me weeks to build and it's real code. My content lab tracks how each of my reels performs against my usual numbers, pulls the outlier posts in my niche, and turns a finished video into six versions with different opening hooks in about 45 seconds, so I can test which hook wins. The editing itself (captions, screenshots that pop in on the point, zooms, music) runs through a Claude skill that edits in my house style.

You don't need any of that to get most of the value this weekend. The pipeline is just three steps, and each one is a prompt or a skill.

Build yours.

1. Ideas. Every week, feed it two sources: your competitor outliers from Build 4, and your own best posts. Ask:

Read the latest report in competitors/reports/ and my notes on my best-performing posts. Give me 10 content ideas for this week in content/ideas.md. For each one: the hook, the one idea it delivers, and which outlier or past post it's based on. Only ideas that fit what we sell (context/work.md).

1. Script. Turn your scripting rules into a skill, so you're not re-pasting them every time. A skill is a folder with a SKILL.md file inside. In Claude Code it lives at .claude/skills/reel-script/SKILL.md. On claude.ai you zip the folder and upload it under Customize > Skills (turn on code execution first, under Settings > Capabilities). Here's a starting point:

---
name: reel-script
description: Use when I ask for a reel, short or TikTok script, or hand you a video idea to script.
---

Before writing, read context/writing-style.md and context/work.md.

1. Give me 3 hook options. Each one names the payoff or makes a claim
   in the first two seconds.
2. Write it to be spoken out loud: short words, contractions, no
   lists read off a screen.
3. Beside each line, add what's on screen while I say it.
4. End with one call to action, nothing else.
5. Keep it under 60 seconds spoken (about 150 words) unless I say otherwise.

The description matters most. It's how Claude decides when to use the skill, so say plainly when it applies.

1. Edit. Two honest options. The easy one: record, then edit in CapCut or Descript, using the on-screen notes from the script as your shot list. The ambitious one: Claude Code can edit video with code (cutting pauses, burning in captions, popping in images), which is how mine works. That's a weekend on its own, and I've written it up below.

Done when you can go from Monday's competitor report to a filmed, edited post in one sitting.

Go deeper: how I make videos with Claude Code, in Making Videos with Claude Code.

How it all fits together

Look at what each build reads from:

Every arrow points back to the first build. That's why it's the foundation, and why skipping it makes the other four feel like party tricks. AI is great, but without your context in place, there's no way it can know you or sound like you.

Where the DIY version stops

This setup runs my day. It works because it's mine, on my laptop, and because I keep it current.

That's also exactly where it stops. It's one person's brain. Your team doesn't get it unless they each build their own, it doesn't update itself when something changes in your inbox or your CRM, and nobody else can trust it as the source of truth for the business.

The version that changes a company is the same idea, built once for everyone. It's one shared brain of how the business works, connected to the tools you already use and kept current on its own, with workflows on top that do the repetitive work: inboxes, follow-ups, reporting, invoicing. That's what we build at Whitehorse, always starting with the brain. If you get this running and want it for your whole business, reply to the welcome email with audit and I'll tell you how we'd start.

Start here

Tonight, not Saturday:

1. Make the folder.
2. Paste the Build 1 interview prompt.
3. Answer five questions honestly.

That's twenty minutes, and it's the foundation everything else stands on. You'll feel the difference in your very next chat.

Built by Alex Sidhu, Whitehorse AI.