What is the recommended learning path?

Follow the numbered flow: (1) How LLMs Work, (2) Prompt Lab, (3) Whiteboard, (4) Website Builder, (5) RAG Explained, (6) MCP & Tool Use, (7) Agent Teams, (8) Agent Library, (9) Idea Factory.

Hard-Won Wisdom

Tips & Lessons — What 4 Years of Daily AI Use Taught Us

Q: What is the difference between an agent, a tool, and a skill?

An agent is an AI system that takes autonomous action. A tool is a single capability like send email or read file. A skill is a multi-step workflow that chains tools together. Tools are atomic; skills are choreography. A good agent has four parts: personality, goals, tools, and skills.

Skip the hype. Here's what actually works — everything we wish someone had told us before we started, and everything we learned the hard way after.

Workshop Tips

🎯01

AI Is Not a Search Engine

Stop asking it questions. Start giving it jobs.

📏02

Be Specific or Be Disappointed

Vague prompts get vague results. Name the format, tone, length, audience.

🖼️03

Show, Don’t Just Tell

One good example beats a hundred words of instruction.

🧩04

Think in Systems, Not Prompts

A prompt is a question. Prompt + tools + goals is an agent.

🚀05

Start Ugly, Ship Fast

Your first version will be rough — that’s the point. Ship, then iterate.

🧠06

Context Is Everything

Background, constraints, examples, audience — feed it all.

🔁07

Let AI Check Its Own Work

Ask it to review and improve its own output. Self-correction is a superpower.

⚙️08

Automate the Boring Stuff

Do something more than twice? Make an agent do it.

👀09

Read the Output, Not Just the Answer

Read its reasoning — you’ll learn how it thinks.

🎭10

One Agent, One Job

One clear role each, then orchestrate them like a team.

🪜11

Break Big Problems Into Steps

AI handles small, clear tasks better than big, vague ones. Chain them.

🎬12

You’re the Director, Not the Coder

You describe and review. AI writes the code. You decide.

What 4 Years of AI Taught Me

🤥01

AI Makes Things Up

With no real data, it invents convincing answers. Verify before you trust.

✋02

AI Will “Improve” What You Didn’t Ask For

Be explicit: “match exactly, do not improve.”

🔍03

Your First Review Will Miss Things

First-pass reviews always miss differences. Budget 2+ rounds before you ship.

📡04

Raw Data Is Not Signal

More data won’t help unless you explain what it means. Context > prompts.

⚠️05

Small Mistakes Cause Big Failures

A typo, a missing import, one wrong character — tiny errors cascade. Sweat details.

🕵️06

Don’t Restart, Investigate

When it breaks, find the root cause or you’ll loop the same failure.

✅07

“Done” Means Proven

If you can’t show it working, it’s not done. Proof is the only definition.

🐛08

Bugs Travel in Packs

Find one problem, look for more. Issues hide behind each other.

👑09

Context Is King

Shorter, focused context = more precise direction. Relevant beats more.

📓10

Keep a Lessons Log

Capture what went wrong and what you learned — or you’ll repeat it.

🎨11

Give Agents Personalities

Role, tone, boundaries shape better output. Give each an identity.

⏰12

Agents That Don’t Learn Are Cron Jobs

Build learning loops in, or they’re scripts with extra steps.

🟢13

Green Doesn’t Always Mean Working

A passing test can lie. Pin tests to frozen fixtures so “green” means something.

Common Mistakes (and What to Do Instead)

We see these patterns in every workshop. Avoid them and you’ll be ahead of 90% of AI users.

Don’t

Use AI as a search engine. “What is machine learning?”

Instead

Give it a job. “Explain machine learning to my sales team in 3 bullet points they can use in client calls.”

Don’t

Copy-paste AI output straight into production without reading it.

Instead

Review every output. AI is a first draft machine, not a final draft machine. You’re the editor.

Don’t

Start by building a complex multi-agent system on day one.

Instead

Start with a single prompt that solves one real problem. Scale to agents after you’ve mastered the basics.

Don’t

Dump your entire knowledge base into the prompt and hope for the best.

Instead

Feed relevant context only. A focused 500-word brief outperforms a 50-page doc dump every time.

Don’t

Send a prompt with no role, no format, and no constraints.

Instead

Always define: who the AI is, what format you want, and what it should NOT do. Constraints improve output.

Don’t

Use the same AI model for every task regardless of complexity.

Instead

Match the model to the task. Fast models for simple work, powerful models for complex reasoning. Save money and time.

Don’t

Expect a perfect result on the first try and give up when it’s not.

Instead

Plan for 2–3 iterations. First pass = rough draft. Second pass = refined. Third pass = polished.

Don’t

Automate a workflow with AI and walk away with no monitoring.

Instead

Set guardrails, add logging, and review outputs regularly. Trust but verify — especially with autonomous agents.

Don’t

Trust a green test suite and move on without looking at what it checks.

Instead

Check what the test actually reads. A test pointed at live, changing data passes while proving nothing — freeze fixtures so “green” means working.

Notice the pattern? Most of these lessons are one idea — the harness: the rules, checks, and memory you build around an agent so it finishes reliably. See the six harness habits on the Agent Teams page →

Peter’s AI Stack

The actual tools used daily to run 300+ autonomous agents across 4 machines. No affiliates — just what works.

🧠

Claude

Primary AI

Anthropic’s model. Best at reasoning, code generation, and long-context work. The brain of everything.

💻

Claude Code

CLI Agent

Command-line AI that reads, writes, and runs code autonomously. Spawns subagents for parallel work.

✨

Gemini

Fallback & Scripts

Google’s multimodal model. Handles script generation, image analysis, and serves as LLM fallback.

⚡

Groq

Fast Inference

Lightning-fast LLM inference for real-time tasks. When speed matters more than depth.

🔌

MCP

Tool Protocol

Model Context Protocol — connects AI to external tools. Like USB for agents. Created by Anthropic.

🗄

Supabase

Vector DB & RAG

Stores embeddings, powers the pRAG chatbot, handles auth. The memory layer for all agents.

🎙

edge-tts

Voice Generation

Microsoft neural TTS. Powers the daily Saarvis Intel video briefings. Free and fast.

🦀

OpenClaw

Agent Framework

Open-source agent orchestration. Manages multi-agent workflows, tool routing, and autonomous task execution.

🎬

Remotion

Programmatic Video

Video built from React code. Narration-driven episodes rendered from templates — the engine behind the agent-narrated shows.

Open Source by Peter

Production-grade tools extracted from the pipeline. MIT licensed. Zero bloat.

🏭

agent-factory

AI Agent Library + Generators

105 AI agent prompts across 12 departments, 20 industry playbooks, AI glossary, and LLM-powered generators that create fresh content automatically. Self-expanding library.

GitHub →

🔄

llm-failover

LLM Failover Library

Multi-provider LLM failover with rate-limit handling and automatic retry. Groq → Gemini → Cerebras. Zero deps.

GitHub →

🔍

seo-audit

SEO Scanner

38-check SEO scanner for static sites. Auto-fix, snapshot diffs, Lighthouse CWV, Google Search Console. Stdlib only.

GitHub →

📊

polydoge-api

Prediction Stats API

Public REST API for PolyDoge prediction stats. 4 endpoints: open picks, resolved history, performance stats, signal analysis. Auto-updated every 30 min.

Live API →

Prompt Cheat Sheet

Copy-paste these templates. They work across Claude, ChatGPT, Gemini — any model. See the full cheat sheet →

The Role PromptDomain-specific expertise

You are a [role] with [X years] of expertise in [domain]. Your audience is [who]. Your tone is [how]. Your goal is to [what]. Now help me with: [your actual request]

The Prompt FixerImprove weak prompts

Here is a prompt I wrote: """ [paste your prompt here] """ Grade this prompt A through F. Then rewrite it to be an A. Explain what you changed and why.

The Agent SpecBuild a custom agent

Design an AI agent with these specs: IDENTITY: [who is this agent? role, personality, expertise] SKILLS: [what can it do? list 3-5 specific capabilities] TOOLS: [what does it have access to? APIs, files, databases] GOALS: [what should it achieve? success criteria] GUARDRAILS: [what should it never do?] Generate a complete system prompt for this agent.

The Chain-of-ThoughtStep-by-step reasoning

I need you to solve this problem step by step. Problem: [describe it] Think through this methodically: 1. First, identify what we know 2. Then, identify what we need to figure out 3. Work through each step, showing your reasoning 4. Check your work before giving the final answer Do not skip steps or jump to conclusions.

The Output ControllerEnforce specific format

Return your response in this exact format: FORMAT: [JSON / markdown table / bullet points / numbered list] FIELDS: [list exactly what to include] LENGTH: [word count or line count] TONE: [professional / casual / technical] DO NOT include: [what to exclude] EXAMPLE: [paste one example of the ideal output]

Frequently Asked Questions

Questions we hear in every workshop.

Will AI replace all jobs?

No — but it will replace people who don’t learn to work with it. That’s not a bumper sticker, it’s what the research actually shows.

A March 2026 study out of Sun Yat-sen University and Alibaba (SWE-CI) tested whether today’s best AI agents could maintain real software codebases over time — not just fix a single bug, but do the ongoing work that human developers do: adding features, adapting to changing requirements, keeping things from breaking.

The result? They break things more than they fix them. Most models had a zero-regression rate below 25%, meaning 3 out of 4 times they touched working code, something else broke. AI is strong at one-shot tasks — answer a question, fix a bug, generate a file — but falls apart when the work requires long-term judgment, architectural thinking, and knowing when not to change something.

This pattern extends beyond software. AI is exceptional at tasks that are repeatable, bounded, and well-defined. It struggles with work that requires context accumulated over time, cross-functional judgment, and accountability for downstream consequences.

The jobs that survive aren’t the ones AI can’t do today — they’re the ones that require sustained ownership of outcomes over time. Architects, not bricklayers. Strategists, not summarizers.

Read the full paper (PDF) · View on arxiv →

Do I need to know how to code?

No. Most of this workshop is about prompting, thinking in systems, and directing AI — not writing code. When we do build things (like in the Website Builder), the AI writes the code for you. You just describe what you want.

Which AI tool should I start with?

Claude (by Anthropic) for reasoning and writing. It handles long documents, writes clean code, and follows complex instructions better than alternatives. Start with the free tier at claude.ai, then explore Claude Code when you’re ready to build agents.

How much does it cost to use AI?

Most tools have free tiers. Claude Pro is $20/month. For the workshop exercises, free tiers are enough. If you’re running agents at scale (like Peter does with 300+), API costs vary — but you can start at $0 and scale up as the ROI proves itself.

What’s the recommended learning path?

Follow the numbered flow on the Start Here page: (1) How LLMs Work — understand the engine, (2) Prompt Lab — practice prompting, (3) Whiteboard — map your ideas, (4) Website Builder — build something real, (5) RAG Explained — fix hallucinations, (6) MCP & Tool Use — connect to the world, (7) Agent Teams — go multi-agent, (8) Agent Library — browse ready-made agents, (9) Idea Factory — practice with challenges.

What is an LLM and how does it actually work?

LLM stands for Large Language Model. It’s a prediction engine: you give it text, it predicts the most likely next word, one token at a time. That’s the entire trick — it just works shockingly well at scale. Try the interactive demo to see tokenization, attention, and prediction in action.

What’s the difference between an agent, a tool, and a skill?

An agent is an AI system that takes autonomous action. A tool is a single capability the agent can call — “send email,” “read file,” “query database.” A skill is a multi-step workflow that chains tools together — “research a topic, draft an article, optimize for SEO, and publish it.” Tools are atomic; skills are choreography. A good agent has four parts: personality, goals, tools, and skills.

What is RAG and when do I need it?

RAG (Retrieval-Augmented Generation) gives AI access to your specific data — documents, databases, knowledge bases — at query time. Use it when you need AI to answer questions about your data, not generic internet knowledge. Without RAG, AI hallucinates confidently. With RAG, it cites sources. See the full explainer.

What is MCP (Model Context Protocol)?

MCP is an open standard (created by Anthropic) that lets AI connect to tools in a universal way. Before MCP, every app built custom integrations. With MCP, you build one server for Gmail and every AI app can use it. Think USB-C for AI. See the interactive explainer.

What is temperature and how should I set it?

Temperature (0.0 to 1.0) controls how creative or random AI output is. Low (0.0–0.3): deterministic, safe, consistent — use for code, data extraction, factual Q&A. Mid (0.3–0.7): balanced variety — use for marketing copy, email drafts, brainstorming. High (0.7–1.0): wild and creative — use for first drafts and idea generation. Try the temperature slider to see the difference.

What is a Personal RAG (pRAG)?

A pRAG is a RAG system built on your knowledge — blog posts, talks, investor memos, project docs. Instead of generic AI answers, you get responses grounded in your actual expertise and body of work. Peter’s Saarvis chatbot is a pRAG. Read how to build yours.

How do I know if an AI output is accurate?

Verify critical claims. AI generates plausible text, not guaranteed facts. Use it like a very fast first draft — always review, fact-check important claims, and test code before shipping. The golden rule: trust but verify. For mission-critical accuracy, use RAG to ground responses in real data.

What’s the difference between an agent and a chatbot?

A chatbot answers questions. An agent takes actions. Chatbots are ping-pong: you ask, it replies. Agents are autonomous: you give them a goal, they plan steps, use tools, evaluate results, and iterate. Agents do work; chatbots give answers.

Can I use AI in my specific industry?

Yes. AI is horizontal — it applies to every industry. Healthcare, finance, legal, education, real estate, manufacturing, creative work. The patterns are the same: automate repetitive tasks, augment decision-making, and build systems that scale your expertise.

How do I connect AI to my tools (email, Slack, databases)?

Through MCP (Model Context Protocol). Install an MCP server for each tool you want AI to access. For example, a Gmail MCP server lets any AI agent read and send email. No custom code needed — the protocol handles the connection. See the MCP & Tool Use page for a visual walkthrough.

Can multiple AI agents work together?

Yes — and they should. An orchestrator agent breaks a task into subtasks and delegates to specialist agents working in parallel. One agent researches, another writes, a third reviews. This is how Peter runs 300+ daily operations. See Agent Teams to learn multi-agent patterns.

How do I build my first agent?

Start with a system prompt. Give it a role, personality, clear goals, and access to 1–2 tools. Use the Agent Teams builder to design one visually. Then try Idea Factory for guided exercises. Don’t over-engineer — ship ugly, iterate fast.

How long does it take to get good at this?

You can write effective prompts in an afternoon. Building simple agents takes a weekend. Running a multi-agent system takes weeks of iteration. The learning curve is fast — the key is to practice daily. Use AI for real work, not toy examples.

AI Glossary

Plain-English definitions of AI terms you'll actually encounter. No PhD required. New terms added weekly.

Browse the Full Glossary →

A Message from Saarvis

The AI agent behind the network has something to say.

Enjoyed these tips?

This free resource took 4+ years of real-world AI work to put together. If it helped you, I'd genuinely appreciate a recommendation on LinkedIn — and feel free to connect while you're there!

If you'd like to leave a tip or donation, consider becoming a Node Supporter ($75) of the Bitcoin Race Team USA. Every bit helps fuel the mission.

Message: Your AI workshop has been really helpful. Made a landing page/portfolio page for all my projects. — Peter gets a lot of these every day <3