Most people start learning AI by hoarding prompt templates and chasing every new model that drops. This roadmap does the opposite, and it’s faster. I came across a video from Jeff Su that lays out exactly how he’d learn AI from scratch in 2026, and it reframed a lot of habits I thought were still useful.
Here’s the big idea from the creator: skip the 80% of AI advice that’s outdated or pure theory, and lock onto the 20% that’s practical and will still matter a decade from now. He organizes it into three levels that stack on top of each other. I’ll break each one down so you can figure out where you actually are right now.
Level 1: Pick one model and go deep
The old way was to bounce between ChatGPT, Claude, and Gemini, always hunting for the “best” one. The original poster argues that’s a waste of energy now. His reasoning is simple: the top models have clustered so close together in capability that the difference for an average user is tiny, and since the big AI companies keep copying each other’s features, the skills you build on one carry straight over to the rest.
So he narrows the field to three real contenders: ChatGPT, Claude, and Google Gemini. Then he gives three principles for choosing:
- Prioritize paid tiers. The gap between free and paid is night and day. If your job hands you paid Gemini, go deep on Gemini even if you started on free ChatGPT.
- Match the model to your work. ChatGPT is the most mature with the most tutorials and strong web search. Claude shines at writing, design, and coding (which quietly powers data analysis and clean diagrams). Gemini wins if you work across text, images, audio, and video, or live inside Google Workspace.
- Go with the vibes. Sounds silly, but the expert’s point is real: the more you enjoy an AI’s personality, the more you’ll use it, and the better you’ll get.
One more move he flags: change your defaults. Companies quietly route you to the weakest model because it’s cheapest for them to run. For real work, always select the most capable model you have access to, since it breaks down your request, maps the steps, and catches nuances you forgot to mention.
Level 2: Context beats the perfect prompt
Here’s where I got genuinely excited. The author makes the case that your prompt is no longer the biggest factor in output quality. Models got powerful enough to infer the role, format, and tone on their own, as long as you hand them a clear outcome and the right context.
His example sold me. Say you need a workout plan. Old way: write a giant prompt spelling out your experience, equipment, schedule, and goals. New way: paste in an article on the push-pull-legs routine as context, then write something short like “build me a 4-day muscle growth routine, 45 minutes a day, based on this.” The second version wins every time, because the AI infers a sharper role from the context than you’d have thought to ask for.
The only framework worth remembering, he says, is OC: Outcome plus Context. And he shares three ways to feed AI the right context:
- Name a real framework. Telling it to “rewrite this using the pyramid principle” packs more context into two words than a paragraph of explaining. You can even ask the AI which frameworks fit your task first.
- Show real examples of good. Paste your last two or three approved status updates, add raw notes, and say “write this week’s in the same format.” Examples carry all the stuff you forget to say out loud.
- Connect your tools. Your best context already lives in Gmail, Drive, Slack, or Notion. Hook those up so the AI pulls files directly instead of you downloading and re-uploading.
Then comes the storage piece. To stop repeating yourself on recurring work, the creator points to Projects in ChatGPT and Claude (called Gems in Gemini). A project holds three things: instructions (your always-on rules), knowledge files (your source docs and examples), and memory (auto-updated milestones). 💡 Pro tip he drops: use markdown (.md) files instead of PDFs, since they’re easier and cheaper for AI to read, and you can ask the AI to convert PDFs for you.
Level 3: Connect projects into a system
Projects are great, but each one is a silo. Your workout coach can’t see your annual health report sitting in another project. This is where the expert introduces the AI system, which does two things projects can’t:
- Pulls context across projects to spot patterns one project would miss. His example: cross-referencing a health checkup with a workout plan, then flagging that there’s no cardio despite borderline high cholesterol.
- Updates itself from your feedback so learnings compound. He calls it the “reconcile” move: you edit the AI’s draft, tell it to reconcile your version with its original, and it proposes rules to remember next time.
He lays out three options by skill level: Gemini Spark (most beginner-friendly, pre-connected to Google tools, less control), Claude Cowork (built for non-technical folks, more control, some setup), and Claude Code / OpenAI Codex (fully customizable power tools if you’re comfortable with code).
Where to start
The honest takeaway from the person who made this: most people aren’t at Level 3 yet, and that’s fine. There’s no rush. Start at Level 1, go deep on one model, then climb when you’re ready.
The full video has the live demos and a couple of HubSpot Gemini tips worth seeing, so check it out and figure out which level you’re on. 🚀