Frequently Asked Questions About Jeff Su AI From Scratch 3-Level System
21 answers covering everything from basics to advanced usage.
// Basics
What does 'skip the useless 80%' actually mean for learning AI?
It means ignoring the vast majority of AI advice — endless tool-hopping, elaborate prompt templates, chasing every new model — and focusing only on the 20% that compounds: mastering one chatbot, supplying strong context, building Projects, and scaling to an AI System. Most AI content optimises for novelty; this system optimises for durable skills that transfer.
What are the three levels in the 3-Level System?
Level 1 is mastering one chatbot with the OC Framework and the most powerful model. Level 2 is building Projects (or Gemini Gems) as permanent homes for recurring work streams using instructions, knowledge files, and memory. Level 3 is migrating multiple Projects into an AI System that cross-references context across silos and self-updates from feedback.
What is the Big Three and why only those three?
The Big Three are ChatGPT, Claude, and Gemini — the only frontier chatbots worth choosing between. All others are non-competitive, derivative, or search tools rather than true frontier models. Since top models have converged in capability and copy each other's features, committing to one of these three gives you skills that transfer to the rest.
Is it worth paying for AI if I only use it occasionally?
If you're building a real workflow, yes — the free-to-paid gap is described as night and day, and the most capable models catch nuances weaker ones miss. Check whether your employer provides paid access first (Paid Tier Priority). If usage is genuinely rare and casual, the free tier is fine, but this system assumes recurring work.
// How To
How do I actually build my first Project step by step?
Pick one recurring work stream. Write Project Instructions covering always-applicable rules and constraints (audience, length, format). Add Knowledge Files — 2-3 approved past outputs plus any named frameworks, all in markdown. Leave Memory for the AI to update automatically. Then run tasks inside the Project so it draws on that context every time without you re-explaining.
How do I use the three context-finding methods?
Method 1, Named Frameworks: name a recognised methodology like 'pyramid principle' or ask the AI which framework fits your goal. Method 2, Real Examples: paste 2-3 approved past outputs instead of describing format in prose. Method 3, Connect Your Tools: point the AI at Drive, Notion, Slack, or email where your context already lives rather than re-uploading files.
How do I run the Reconcile Move correctly?
After the AI produces a draft, edit it to match your voice and standards, then paste your final version back with a prompt like: 'Reconcile my final version with your initial draft and propose rules to remember for next time.' The AI extracts the delta as explicit rules. Do this repeatedly inside a Project so alignment accelerates.
How do I find and switch to the most powerful model?
Open your chatbot's settings or the model selector near the chat input. Companies default you to a lightweight model, so manually choose the highest-tier option your paid plan exposes — for example the most capable reasoning model. Do this at the start of every session until checking becomes automatic habit.
What are the recurring work streams I should turn into Projects?
Look for 2-3 tasks you do repeatedly that need the same context each time — weekly status updates, campaign reports, competitive research, process documentation, or monthly onboarding prep. Anything where you find yourself re-pasting the same background or re-explaining the same rules is a strong candidate for a permanent Project home.
// Troubleshooting
Why is the AI ignoring my Project instructions?
Usually because instructions are too vague, buried in prose, or contradicted by weak context. Rewrite instructions as clear, always-applicable rules and constraints. Replace descriptions of good output with actual real examples in your Knowledge Files. Confirm you're working inside the Project, not a fresh chat, and that you selected the most capable model.
My AI outputs still don't match my voice after several tries. What's wrong?
You're likely skipping the Reconcile Move. Describing your voice in prose rarely works; the system needs to learn from the delta between its drafts and your edits. Paste your edited final version back and ask it to extract rules. Also add 2-3 real examples of your best writing as Knowledge Files so it has concrete evidence.
My knowledge files aren't being read well. How do I fix it?
Convert them to markdown. PDFs are harder for the AI to parse and more expensive in tokens, so tables and formatting often get garbled. Ask the AI to convert each PDF to a clean .md file, then re-upload the markdown version. Keep files focused — a few strong examples beat a large dump of loosely relevant documents.
I'm overwhelmed by too many AI tools. How do I stop tool-hopping?
Commit to one of the Big Three using the three selection principles: paid access first, then work-type match, then vibes. Skills do not compound when you split focus across tools, and top models have converged so depth on one transfers to all. Delete the temptation — give one chatbot a full month before considering anything else.
// Comparisons
How does this system compare to using ChatGPT casually with one-off prompts?
Casual one-off prompting resets context every session, so you repeat yourself endlessly and outputs never improve. This system stores context in Projects and teaches the AI your preferences through the Reconcile Move, so quality compounds. The difference is between a tool you re-explain daily and a system that learns and needs less instruction over time.
How is an AI System different from a Project?
A Project is a single permanent home for one work stream — siloed, so it can't reference other Projects. An AI System connects multiple Projects, enabling two things Projects can't: cross-project context synthesis (spotting patterns across silos) and self-updating rules from feedback. Move to an AI System only once you have several active Projects.
How does the OC Framework compare to CRAFT or RTF prompting templates?
Multi-part templates like CRAFT or RTF spell out role, format, and tone in every prompt — necessary for older, weaker models. The OC Framework drops all of that because modern models infer those elements from context. You only need Outcome plus Context. It's less to remember and produces better results by shifting effort from prompt structure to context quality.
Gemini Spark vs Claude Cowork vs Claude Code — which AI system tier should I pick?
Pick by technical comfort. Gemini Spark suits non-technical users wanting minimal setup — it's pre-connected to Gmail, Calendar, and Drive but offers least control. Claude Cowork suits non-technical users wanting more control with some setup and no coding. Claude Code or OpenAI Codex suit those comfortable with code — fully customisable, maximum power. The model selector UI signals the tier.
// Advanced
How do I make learnings compound across an AI System over time?
Consistently run the Reconcile Move so the system extracts rules from your edits, and give explicit feedback when outputs miss the mark. In an AI System, these corrections update its own rules automatically and apply across connected Projects. The more you use and correct it, the less instruction it needs — that's the compounding effect.
Can I use named frameworks I don't know the name of?
Yes — ask the AI directly: 'What are the best frameworks for [goal]?' then pick one and name it in your prompt. Two words of a framework name carry more context than a paragraph of description, so naming a methodology is one of the highest-leverage context moves available even when you're learning the frameworks yourself.
How should a developer configure the highest AI system tier?
Use Claude Code or OpenAI Codex, which expose all model options in the selector. Connect your existing tools and data sources rather than re-uploading files, migrate your Projects in as structured contexts, and let the system cross-reference them. Keep knowledge files as markdown for efficiency, and rely on the Reconcile Move plus explicit feedback to drive self-updating rules.
How many Projects should I have before building an AI System?
Wait until you have multiple active Projects — typically 2-3 recurring work streams that would genuinely benefit from cross-referencing each other. Building an AI System with only one Project adds complexity without payoff, since the whole value is synthesising context across silos and compounding feedback across connected work streams.