GOOD VIBES

Where we're going

tonight/
├── 00_hello/         prove it's real
├── 01_the_machine/   what Claude actually is
├── 02_the_files/     context, the whole trick
├── 03_two_doors/     Cowork and Code
├── 04_pick_a_brain/  models, tokens, what it costs
└── 05_the_build/     a bot and a digest, yours to keep

Two builds that keep working after you go home: a Telegram bot so your AI can reach your phone, and a morning news digest that writes itself while you sleep.

The second one arrives through the first. That's the trick.

Before we start

You should already have:

• The Claude desktop app, signed in (Pro plan or above)
• Telegram on your phone, account made at home
• An empty folder somewhere obvious, like Documents/claude-workshop

Missing one? Grab a helper now, not at build time.

Your first prompt

Open your workshop folder in the Claude app, then type:

Create a file called hello.md in this folder that introduces
yourself and lists 3 fun things we could build today.

Now watch the folder. A real file appears on your real computer.

A chatbot answers you. This thing does the work.

It's a guessing game

Same machine. One extra sentence. Different answer.

Hold that thought.

How it learned

It read most of the internet with one job: guess the next word. Wrong guess, nudge the dials, go again. Billions of times.

Then people rated its answers until being helpful scored better than merely sounding plausible.

That's the training. There's no database of facts in there, just very good guessing shaped by everything it read.

From guesser to worker

A guesser becomes a worker when you give it a toolbox and a loop.

   read your request
        │
        ▼
   pick a tool ──── open a file, write a file,
        │           run a command, fetch a page
        ▼
   look at what happened
        │
        ▼
   done? ──no──▶ go again
        │
       yes

That's the whole machine. A fast worker with a small toolbox, and you can watch every move it makes.

It's all just files

Everything Claude "knows" is a text file you can open and read.

your-folder/
├── CLAUDE.md      your notes to Claude: who you are, how you like things done
├── skills/        recipes for repeatable jobs. type / to run one
└── memory/        what it keeps between conversations

Want to know why it did something? Open the file and look. Change the file, change the behaviour.

Proof: I run my life on this

dwe_os/
├── CLAUDE.md        house rules: who I am, what lives where
├── notes/           everything I know, in plain markdown
├── work/homenet/
│   └── CLAUDE.md    this client only: their site, their rules
├── personal/code/   29 weekend builds, many with their own CLAUDE.md
└── memory/          what Claude remembers about me between chats

One CLAUDE.md at the top for the house rules. Any folder can carry its own for that project's specifics. Claude reads every one on the path to where it's working: general first, local on top.

Every morning it briefs me: today's meetings, the inbox, last night's training. Ask it "what did we decide about X" and it answers from my own meeting notes.

I never introduce myself twice.

Context compounds

Remember the cat on the keyboard? One sentence of context changed the guess.

Now imagine a folder full of sentences about you. Your projects, your preferences, the way you like things done. Every conversation starts from there.

The more you write down, the less you explain. Your CLAUDE.md starts tonight.

/clear and /compact

The chat window is working memory, and it fills up. The longer a conversation runs, the more old chat Claude re-reads every turn, and the fuzzier it gets.

• /clearWipe the chat and start fresh. Files, CLAUDE.md and memory all survive. Use it every time you switch jobs.
• /compactSquash the chat into a short summary and keep going. Use it mid-job when the conversation is long but you still need the thread.

New job, /clear. Same job, long chat, /compact.

That's why the files matter: anything worth keeping lives in a file, so wiping the chat costs you nothing.

Two front doors

Same brain, different bench. Tonight we're building, so we use Claude Code, inside the desktop app. No terminal required.

Pick a brain

Claude comes in sizes. Same family, different depth of thought.

• haikuthe quick onesimple jobs, instant answers1x
• sonnetthe workhorsemost building, most days3x
• opusthe deep onehard problems, long jobs5x
• fablethe flagshipwhen nothing else will do10x

Those multipliers are real, straight off Anthropic's price list. On a subscription you never see a bill, you see a usage limit. Same maths: the flagship burns it ten times faster than the quick one.

Start with Sonnet. Reach up when it struggles, down when it's routine. You can switch mid-conversation.

The meter runs on tokens

Claude doesn't read words. It reads tokens, chunks of about three quarters of a word. Remember the guessing machine? Each guess it makes is one token.

The meter counts both directions: what you send in, what it writes out. Words out cost about five times words in.

And the catch: Claude has no memory between turns. Every message re-sends the whole conversation, files and all. Long chats compound.

• Start small. Reach up only when it struggles.
• New job, new chat. Don't make it re-read an hour of history.
• Keep CLAUDE.md tight. Every line rides along on every turn.

Build 1: your AI gets your number

Why Telegram? It's the only big messenger where anyone can make a bot, free, in about 30 seconds.

1. In Telegram, message @BotFather and send /newbot
2. Give it a name and a username
3. Copy the token it hands you. Treat it like a password.
4. Send your new bot /start. It can't message you until you do.
5. Paste this into Claude:

Here is my Telegram bot token: [PASTE TOKEN]
Send a test message to my phone so I know it works.

Nobody tells Claude how. Watch it work the API out, step by step. Then your phone pings.

You just built a connector

That token you pasted wired Claude into Telegram. That's a connector: a plug between Claude and an app out in the world.

The plug standard is called MCP. There are connectors for calendars, email, notes, music, smart homes. The Claude app has a directory of them built in, under Customize.

Remember the toolbox? Every connector you add puts more tools in it. Same loop, longer reach.

Make it yours

Connect something you actually use. Customize, then Connectors: calendar and email are the good first picks.

Then ask for a summary, sent to your pocket:

• "Text my bot my meetings for tomorrow"
• "Summarise my unread email and send it to Telegram"
• "Check my calendar and email, and tell me what tomorrow looks like"

Your real life, landing in the bot you built ten minutes ago.

Connect only what you're happy for Claude to read. Connector not cooperating? Skip it: "send my phone one fact about my hometown I probably don't know".

Meet cron

Before your digest can build itself, it needs an alarm clock. Computers have used the same one since the 1970s: cron. Five fields that name a moment.

"every morning at 8am"

 0  8  *  *  *
 │  │  │  │  └─ day of week (* = any)
 │  │  │  └─── month
 │  │  └───── day of month
 │  └──────── hour (8)
 └─────────── minute (0)

"every friday at 5pm"     0 17 * * 5
"9am on the 1st"          0 9 1 * *

You'll never have to write one. Say the schedule in English and Claude translates. The routine it saves is just a prompt plus a cron string, parked in the cloud.

When the clock hits, a fresh Claude wakes up, follows the prompt, sends the result, and goes back to sleep. Laptop open or shut, it makes no difference.

Build 2: the daily digest

One prompt. Pick your city, pick your news.

Create a scheduled routine that runs every morning at 8am
and messages my Telegram bot with:
1. Today's weather in [YOUR CITY] (use wttr.in, no key needed)
2. The top 3 headlines from Hacker News, one line each
3. One interesting "on this day in history" fact
Keep it short enough to read while the kettle boils.
Run a test now so I can see it.

Why this clicks

The digest posts to the bot you built twenty minutes ago. Two builds, one machine.

It runs on a schedule, in the cloud, with your laptop shut. That's the Cowork trick from earlier, and you just wired it yourself.

Tomorrow at 8am, it proves it.

live demo: my actual morning brief