BUILD YOUR FIRST AI EMPLOYEE.

@pro.glitch

85.6K followers · usually 2.6K views a post

Views

39.6K

Likes

1.3K

Comments

732

SoundTikTok sound · original sound - pro.glitch

View velocity

+17.8K views over 15.4d · ~1.2K/day · 4 observations

Promoted product

Professor Glitchaskglitch.com

Learn to build with AI — not just chat with it. Professor Glitch teaches founders, creators, and operators how to design AI systems that ship real work: agents, automations, and workflows you actually use.

11 videos

Why it worked

The video breaks down a complex topic into digestible, actionable steps, using clear visuals and a confident, engaging presenter. The "build it yourself" approach combined with the promise of a "real AI employee" that runs on its own is highly appealing to entrepreneurs and tech enthusiasts.

Summary

The video outlines a 7-step process for building an AI employee, emphasizing the importance of a loop, specific tools, context, guardrails, and continuous correction. It contrasts a functional AI employee with a "toy" that lacks these elements and encourages viewers to learn the process from scratch via a link in the creator's profile.

Structure

1Build your first AI employee.
2An agent is a loop.
3

Original caption

The exact roadmap to build your first AI employee — from zero (the whole build) The exact roadmap to build your first AI employee — from zero. One that works around the clock, does a real job, and gets sharper every week because it learns from everything you do. Built on how the best actually do this (Anthropic literally published the rules) + running my own every day. Save this — here's the whole build: First, what an "agent" even is: a LOOP (goal → step → check → next, until it's done) that can reach for TOOLS. That's it. But a raw agent hits a wall — it forgets everything, only moves when you poke it, owns nothing, and nobody notices when it breaks. A brilliant temp with amnesia. The roadmap turns that into staff: PICK ONE JOB → small, boring, repeatable. Not "run my business." Rule one from the pros: start stupidly simple. Don't build complex until simple has failed you. WRITE THE INSTRUCTIONS → do the task once by hand in a chat, narrating as you go. That walkthrough IS the file. Your employee is only as good as what's written in it. GIVE IT TOOLS → let it reach your inbox / sheet / CRM and actually DO the job. Keep it few and clearly labeled — 30 tools confuse it, 3 sharp ones don't. THE RIGHT CONTEXT, NOT ALL OF IT → real research calls it "context rot": the more you cram in, the worse it gets at finding what matters. Feed it only the facts this one job needs. Knowing what to leave OUT is the skill. GUARDRAILS → decide what it does alone vs what waits for your yes. Small reversible stuff runs solo; anything that spends money or can't be undone STOPS and asks you. (The #1 way these blow up = no human check on the irreversible move.) PUT IT ON A TRIGGER → 9am, a new lead, every Friday. The moment it runs on its own, it's an employee, not a tool you open. THE LOOP → when it's wrong, you correct it like a person; the fix gets written back into the instructions, so it never makes that mistake twice. That's how it compounds — a little sharper every week. Skipping this is why most people's AI quietly gets WORSE. Build one right and you don't start over — you copy the pattern. One becomes a team. A team becomes a business that runs while you sleep. Want to learn all of it from scratch — AI, automations, agents, AI employees, the whole thing, deep? Check the link in my profile. #aiagents #professorglitch

More from @pro.glitch

On-screen text

01 BUILD YOUR FIRST AI EMPLOYEE. WORKS AROUND THE CLOCK LEARNS FROM EVERYTHING YOU DO NOT A CHATBOT. NOT A ONE-OFF. HIRE. SAVE THIS ONE LONGER THAN USUAL. 02 ONE BECOMES A TEAM. EVERYBODY'S ABOUT TO BUILD THESE GROUNDED IN THE REAL RULES the people who make Claude published the rules I run my own employees every single day ONE - A TEAM not the surface. THE WHOLE BUILD. "Let's go." RUNNING MY OWN, EVERY DAY 03 START AT THE BOTTOM AN AGENT IS A LOOP. GOAL -> STEP -> CHECK -> NEXT A LOOP. checks itself, every single pass plus one more thing... 04 TOOLS = AN AGENT. REACH OUTSIDE ITSELF NOT JUST TALKING PICK UP A TOOL. HIT AN APP. YOUR HAND A TOOL = AGENT REAL HANDS, NOT JUST WORDS email M calendar you've used one. THAT'S AN AGENT. YOU'VE USED ONE 05 THE WALL. IT HITS FAST FORGETS EVERYTHING WAITS FOR YOU OWNS NOTHING NOBODY NOTICES a very smart temp with... AMNESIA NO SCHEDULE NO JOB WE FIX ALL FOUR. ONE AT A TIME 06 STEP 1 - PICK ONE JOB STEP 1 / 7 1 STEP NOT "RUN MY BUSINESS." ONE TASK. PICK ONE JOB. WHAT COULD IT BE? start stupidly simple. SMALL, BORING, REPEATABLE. ✓ small ✓ boring ✓ repeatable NOT "RUN MY BUSINESS" ANTHROPIC'S OWN RULE don't build complicated until simple has failed you. DO ONE THING SO WELL - YOU FORGET YOU EVER DID IT. 07 STEP 2 - INSTRUCTIONS STEP 2 / 7 2 STEP DO IT ONCE BY HAND. NARRATE AS YOU GO. WRITE THE INSTRUCTIONS. THE TRICK THAT SAVES YOU HOURS DO IT ONCE -> SKILL.MD = THE FILE. real data, real steps narrate as you go = THE INSTRUCTION FILE. THAT FILE IS EVERYTHING only as good as what's written in it. 08 STEP 3 - TOOLS STEP 3 / 7 3 STEP THE HANDS, NOT JUST WORDS GIVE IT TOOLS. REACH INTO YOUR REAL APPS 30 TOOLS 3 TOOLS your inbox M your sheet your CRM FEWER, SHARPER TOOLS. WIN EVERY TIME. 09 STEP 4 - RIGHT CONTEXT STEP 4 / 7 4 STEP NOT ALL OF IT. THE RIGHT CONTEXT. EVERYTHING YOU KNOW -> 3 CLEAN FACTS CUSTOMER YOUR RULES YOUR TONE ALMOST EVERYONE BLOWS THIS. the instinct is to pour everything in. DON'T. A GIANT MEMORY = SMART it's a confused employee customer, rules, tone - that's it 10 STEP 4 - CONTEXT ROT STEP 4 / 7 4 STEP KNOWING WHAT TO LEAVE OUT CONTEXT ROT. MORE CONTEXT = WORSE RECALL CUSTOMER YOUR RULES YOUR TONE A GIANT MEMORY = SMART it's a confused employee customer, rules, tone - that's it ALMOST EVERYONE BLOWS THIS. LEAVING IT OUT IS THE REAL SKILL. 11 STEP 5 - GUARDRAILS STEP 5 / 7 5 STEP NEVER SKIP THIS ONE SET THE GUARDRAILS. SMALL + REVERSIBLE -> RUNS ITSELF DOES IT ALONE WAITS FOR YOUR YES small + reversible -> runs itself. YOU DECIDE, UP FRONT what it does alone, and what waits for your yes. small stuff? it just runs. 12 STEP 5 - YOUR YES STEP 5 / 7 5 STEP NO HUMAN CHECK = THE #1 BLOWUP YOUR YES, FIRST. SPENDS MONEY. SENDS BIG. CAN'T UNDO. SPENDS MONEY THE BIG SEND CAN'T UNDO YOUR YES PUT THAT CHECK IN. HAND IT THE REST. 13 STEP 6 - TRIGGER STEP 6 / 7 6 STEP A TOOL YOU OPEN - STAFF THAT RUNS PUT IT ON A TRIGGER. IT SHOWS UP ON ITS OWN 9:00 AM 4 A LEAD COMES IN EVERY FRIDAY you stop poking it. A TOOL YOU OPEN -> STAFF THAT RUNS. 9:00 AM a lead comes in every Friday you stop poking it. 14 STEP 7 - THE LOOP STEP 7 / 7 7 STEP THE ONE ALMOST NOBODY DOES THE LOOP. WRONG -> CORRECTED -> IT STICKS GETS IT WRONG YOU CORRECT IT IT STICKS THE PAYOFF correct it right there, like you'd correct a person. IT TRAINS ON EVERYTHING YOU DO. 15 STEP 7 - SCOREBOARD STEP 7 / 7 7 STEP SAME MISTAKE TWICE? instead of a repeat if the answer's no - it's working. it's compounding. EVERY MISTAKE - A PERMANENT LESSON SAME MISTAKE TWICE? NO instead of a repeat if the answer's no - it's working. it's compounding. SO DON'T SKIP IT. 16 THE EMPLOYEE 16 A REAL AI EMPLOYEE. ONE JOB INSTRUCTIONS TOOLS RIGHT CONTEXT YOUR CONTROL A TRIGGER THE LOOP a boundary you control a trigger the loop BETTER EVERY WEEK. 17 THE TEAM 17 COPY THE PATTERN. BUILD ONCE - COPY THE PATTERN RUNS WHILE YOU SLEEP. 2ND = HALF THE TIME 3RD = A FRACTION you don't start over for the next one. YOU COPY THE PATTERN. A BUSINESS THAT RUNS WHILE YOU SLEEP. 18 THE HONEST PART. 18 THE STEPS EVERYONE SHIPS never corrected -> never improves context discipline the guardrails the loop NOT A MAGIC BUTTON. these decide whether yours is real staff or a toy. an employee you never correct, never improves. DO THEM 19 LEARN IT FROM SCRATCH. 19 LEARN IT FROM SCRATCH. AI AUTOMATIONS AGENTS AI EMPLOYEES IF THIS SOUNDS FOREIGN but you actually want to learn it HOURS OF CONTENT. FROM THE GROUND UP. SERIOUS ABOUT IT? check the link in my profile.

Transcript

I'm going to show you the exact roadmap to build your first AI employee — from zero. One that works around the clock, does a real job, and gets sharper every week because it learns from everything you do. Built on how the best actually do this (Anthropic literally published the rules) + running my own every day. Save this — here's the whole build: First, what an "agent" even is: a LOOP (goal → step → check → next, until it's done) that ca Start at the bottom. An agent is a loop. Goal, step, check, next. A loop. Checks itself, every single pass. Plus one more thing... Tools = an agent. Reach outside itself. Not just talking. Pick up a tool. Hit an app. Your hand. A tool. = Agent. Real hands, not just words. Email, Gmail. Calendar. You've used one. That's an agent. You've used one. The wall. It hits fast. Forgets everything. Waits for you. Owns nothing. Nobody notices. A very smart temp with... Amnesia. No schedule. No job. We fix all four. One at a time. Step 1 - Pick One Job. Step 1 / 7. 1 Step. Not "run my business." One task. Pick one job. What could it be? Start stupidly simple. Small, boring, repeatable. Small. Boring. Repeatable. Not "run my business." Anthropic's own rule: Don't build complicated until simple has failed you. Do one thing so well - you forget you ever did it. Your first employee should do one thing so well - you forget you ever did it yourself. Step 2 - Instructions. Step 2 / 7. 2 Step. Do it once by hand. Narrate as you go. Write the instructions. The trick that saves you hours. Do it once. -> Skill.MD = The File. The instruction file. That file is everything. Only as good as what's written in it. Step 3 - Tools. Step 3 / 7. 3 Step. The hands, not just words. Give it tools. Reach into your real apps. 30 tools. 3 tools. Your inbox. Your sheet. Your CRM. Fewer, sharper tools. Win every time. Step 4 - Right Context. Step 4 / 7. 4 Step. Not all of it. The right context. Everything you know -> 3 clean facts. Customer. Your rules. Your tone. Almost everyone blows this. The instinct is to pour everything in. Don't. There's hard research on this. It's called context rot. The more you cram in, the worse the AI gets at finding what actually matters. Your instinct is to pour everything in. Don't. An employee you never correct, never improves. Step 4 - Context Rot. Step 4 / 7. 4 Step. Knowing what to leave out. Context rot. More context = worse recall. Customer. Your rules. Your tone. A giant memory = smart. It's a confused employee. Customer, rules, tone - that's it. Leaving it out is the real skill. Step 5 - Guardrails. Step 5 / 7. 5 Step. Never skip this one. Set the guardrails. Small + reversible -> runs itself. Does it alone. Waits for your yes. Small + reversible -> runs itself. You decide, up front. What it does alone, and what waits for your yes. Small stuff? It just runs. Step 5 - Your Yes. Step 5 / 7. 5 Step. No human check = the #1 blowup. Your yes, first. Spends money. Sends big. Can't undo. Spends money. The big send. Can't undo. Your yes. Put that check in. Hand it the rest. You can trust it with everything else. Step 6 - Trigger. Step 6 / 7. 6 Step. A tool you open - staff that runs. Put it on a trigger. It shows up on its own. 9:00 AM. A lead comes in. Every Friday. You stop poking it. A tool you open -> staff that runs. 9:00 AM. A lead comes in. Every Friday. You stop poking it. It shows up on its own. 9:00 AM. A lead comes in. Every Friday. You stop poking it. Step 7 - The Loop. Step 7 / 7. 7 Step. The one almost nobody does. The loop. Wrong -> corrected -> it sticks. Gets it wrong. You correct it. It sticks. The payoff. Correct it right there, like you'd correct a person. It trains on everything you do. Correct it right there, like you'd correct a person. And that correction gets written back into its instructions. Next time, it's already fixed. That's how it trains on everything you do. Step 7 - Scoreboard. Step 7 / 7. 7 Step. Same mistake twice? Skipping this = quietly gets worse. Instead of a repeat. If the answer's no - it's working. It's compounding. Every mistake - a permanent lesson. Same mistake twice? No. Instead of a repeat. If the answer's no - it's working. It's compounding. A little sharper every single week. Skipping that one check is the number one reason people's AI quietly gets worse instead of better. So don't skip it. 16. A real AI employee. One job. Instructions. Tools. Right context. Your control. A trigger. The loop. A boundary you control. A trigger. The loop. Build that once. Copy the pattern. Runs while you sleep. 2nd = half the time. 3rd = a fraction. You don't start over for the next one. You copy the pattern. A business that runs while you sleep. The thing everyone's chasing. Built one employee at a time. 18. The honest part. The steps everyone skips. Never corrected -> never improves. Context discipline. The guardrails. The loop. Not a magic button. These decide whether yours is real staff or a toy. An employee you never correct, never improves. Do them. 19. Learn it from scratch. Learn it from scratch. AI. Automations. Agents. AI Employees. If this sounds foreign, but you actually want to learn it. AI from scratch. Automations. Agents. AI employees. This whole world on a deep level. I've got hours of content from the ground up. If you're even a little serious about it, check the link in my profile.