How & Goals

Stop asking AI questions.
Start giving it goals.

Most people use AI like a search engine: type a question, get an answer, close the tab. That's not how this works. You have an idea. You tell your agent the goal. It builds the process. You have a completed goal. That's the shift.

The Framework

Four steps. Every time.

1

You have an idea

It starts in your head. A business problem. A creative project. A workflow that's eating your time. Something you want to exist that doesn't yet.

2

You tell your agent the goal

Not "how do I do X?" — that's a question. Instead: "I want X to exist by Friday." The difference is everything. Questions get answers. Goals get action plans.

3

AI generates the process

Your agent breaks the goal into steps. It identifies what it can do autonomously and what needs your input. It starts executing. It asks you to decide only when it genuinely can't.

4

You have a completed goal

Not an answer to read. Not a suggestion to evaluate. A finished thing that exists in the world because you stated what you wanted and your agent built it.

This is the mindset shift

When you sit down with ChatGPT or Gemini or Claude, you're having a conversation. It's useful. It's sometimes impressive. But when you close the tab, nothing has changed in your life except that you read some text.

A goal-based agent is different. You say "I want a weekly newsletter that goes out every Monday with AI industry news and a personal take." The agent doesn't just tell you how to do that. It drafts the newsletter. It schedules it. It sets up the delivery pipeline. Next Monday, your subscribers get an email. You didn't write it from scratch — you approved it.

That's not a better chatbot. That's a fundamentally different relationship with technology.

❌ Question

"What stocks did Congress trade this week?"

✅ Goal

"Monitor every congressional trade and alert me when someone buys more than $100K in a stock I own."

❌ Question

"How do I write a LinkedIn post?"

✅ Goal

"Draft three LinkedIn posts this week in my voice, schedule them, and show me for approval."

❌ Question

"What's the weather in Kona?"

✅ Goal

"Track weather at all my upcoming travel destinations and tell me if I need to pack differently."

The How

Set up your own AI agent in 30 minutes.

This is the same stack that FRED runs on. Open source, self-hosted, and built for people who care about owning their data. You don't need to be a developer. You need a computer, a terminal, and the willingness to follow steps.

Step 1

Install the prerequisites

You need two things on your machine before anything else:

Node.js (version 22 or higher)

Node.js is the runtime that powers OpenClaw. It's free, open source, and runs on Mac, Windows, and Linux.

Mac (recommended):

Install Homebrew first if you don't have it, then:

brew install node
Windows / Linux / Manual:

Download from nodejs.org — grab the LTS version.

Verify it worked: node --version should show v22 or higher.

An AI provider API key

Your agent needs a brain. Pick one to start — you can always add more later.

  • Anthropic (Claude) — What FRED runs on. Best reasoning. Best safety. Start here if unsure.
  • OpenAI (GPT) — The most recognizable name. Great all-rounder.
  • Google (Gemini) — Generous free tier. Strong at document processing.

Create an account, generate an API key, and save it somewhere safe. You'll need it in a minute.

Step 2

Install OpenClaw

One command. That's it.

npm install -g openclaw

This installs the OpenClaw gateway globally on your machine. It's the brain that connects your AI models, messaging channels, memory files, and tools into one unified system.

Verify: openclaw --version

Step 3

Configure your agent

Run the setup wizard:

openclaw init

This creates your config file and walks you through the basics: which AI model to use, where to store your agent's memory, and how to connect a messaging channel.

At minimum, you'll need to paste in your API key from Step 1. The wizard handles the rest.

Step 4

Connect a messaging channel

This is what makes it an agent instead of a chatbot. Pick the app you already use:

  • Telegram — Create a bot via @BotFather, paste the token into your config
  • Discord — Create a bot in the Developer Portal, add the token
  • WhatsApp — Connect via WhatsApp Business API or bridge
  • Signal, Slack, iMessage — All supported. Check the docs for setup guides.

Once connected, your agent lives in your pocket. Text it like you'd text a friend.

Step 5

Start your agent

openclaw gateway start

Your agent is now live. Send it a message. Give it a goal. Watch it work.

From here, you can add memory files, set up scheduled tasks, connect your email and calendar, enable browser automation, and build out everything described on this site. The OpenClaw documentation covers it all.

How The Connection Works

Where your data actually goes.

This is the question everyone asks second and should ask first. When your agent uses a frontier model like Claude Opus, GPT, or Gemini, does your hard drive get uploaded to an AI company? Does the model get installed on your machine?

Neither. And understanding why is the whole ballgame.

Think of it as a phone call with a consultant who has no memory

The consultant never enters your building. No key, no badge, no network access, never touches the filing cabinet. You are the one in the office with the files. When you need help on a document, you read the relevant paragraph aloud over the phone. The consultant never sees the document — only the words you chose to say.

And when the call ends, they remember nothing. Next call, you start over.

One turn, mechanically

  1. 1
    Your message hits software running on your own machine. The agent runtime is local. It lives on your hardware, holds your file access, and owns every credential.
  2. 2
    It packages up text and sends it over an encrypted connection. Instructions, the conversation so far, and your new message — as text, over TLS, to the model provider's API.
  3. 3
    The model replies with either an answer or a request. Not an action — a request. "Read this file." "Run this command." "Search the web."
  4. 4
    That request comes back and executes locally. This is the control point. The model has no network path into your machine. It cannot reach in and take anything. It can only ask, and your local runtime decides whether to comply.
  5. 5
    The result goes back up as text on the next turn. And the loop repeats until the goal is done.

The API is stateless — and that surprises people

The model has zero persistent memory of you between calls. Every single turn re-transmits the entire conversation, because otherwise the model would have no idea what you were talking about. There is no account file on their side accumulating your history. All continuity lives in files on your own disk. That's precisely why a real agent keeps memory files locally — it's not a nice-to-have, it's the only way continuity can exist at all.

What leaves, and what never does

Never leaves your machine

  • Every file the agent doesn't open
  • Filenames in directories it never lists
  • Credentials and API keys in your environment
  • Anything outside the current task's scope
  • The overwhelming majority of your drive

Does leave your machine

  • The text of your conversations
  • Contents of files the agent actually opens — in the excerpt it reads
  • Outputs of commands it runs
  • All of it re-sent every turn, so long sessions retransmit repeatedly

The three things vendors won't volunteer

1. "End-to-end encryption" is impossible here

TLS protects data in transit. But the provider's servers must see plaintext to actually reason about it — that's the entire function. Any AI product claiming end-to-end encryption is either confused or lying, and it's a useful tell about how carefully the rest of their security page was written.

2. Fallback chains change who's processing your data

Most serious setups fail over to a second or third provider when the primary is down. That means a request you assumed went to one company went to a different one, under different terms. If you handle confidential material, know your fallback chain and disclose it. It's the single easiest thing to forget.

3. The real attack surface is prompt injection, not eavesdropping

Nobody is intercepting your TLS session. The actual risk is a malicious instruction hidden inside a web page, email, or PDF your agent reads — text engineered to make the agent act against you. Treat all external content as data, never as instructions. Any agent with real access needs an explicit defense here, and you should ask every AI vendor what theirs is.

And when nothing can leave at all

For genuinely privileged material there's a third option: run the model entirely on your own hardware. Tools like Ollama let a capable machine host the model locally, so nothing crosses the internet — zero API calls, zero third-party processing.

The tradeoff is real. Local models are meaningfully weaker than frontier models. So the defensible architecture isn't all-or-nothing, it's tiered: local models for anything confidential, frontier APIs for research, drafting, and general work. That's an answer you can give a client with a straight face.

Model providers' commercial API terms generally state that inputs and outputs aren't used for training by default — but "generally" and "by default" are doing work in that sentence. Read your specific tier. Know your retention window. And be precise about the difference between a well-configured system and a certified one.

What Happens Next

Your first goal should be small.

Don't try to build the whole system on day one. Start with something concrete:

📧

"Check my email every morning and summarize anything urgent."

Connect your email, set up a scheduled task, and wake up to a briefing instead of inbox dread.

📝

"Draft a blog post about [topic] in my voice."

Give your agent a writing sample. Tell it the topic. Review what it produces. Iterate.

📊

"Track these 5 stocks and tell me if anything moves more than 3%."

Connect a market data API, set alert thresholds, and let it watch while you work.

🔒

"Run a security check on my machine every night."

Automated system audits, update monitoring, and breach alerts — all while you sleep.

Each completed goal teaches you something. Each one makes the next one faster. Within a month, you'll have an agent that knows your preferences, your schedule, your writing voice, and your priorities — because you taught it through goals, not prompts.

Ready to build yours?

The AI Agent Playbook covers the full build — from first install to a production agent that runs 24/7. Or skip the reading and book a session where we build it together live.

Get the Playbook — $39 Book a 1:1 Session — $500