Jeff Su AI From Scratch 3-Level System

Last updated: 18 July 2026

Apply a structured three-level methodology to go from AI beginner to compound-learning AI system user, skipping the irrelevant 80% and focusing only on practical, durable skills.

// TL;DR

The Jeff Su AI From Scratch 3-Level System is a structured methodology for going from AI beginner to a compound-learning AI system user by skipping the irrelevant 80% of advice. It moves you through three levels: mastering one chatbot with the OC Framework (Outcome + Context), building Projects as permanent homes for recurring work, then migrating to an AI System that cross-references context and self-updates. Use it when you want to build a sustainable AI workflow from scratch, upgrade ad-hoc AI usage into a systematic approach, or decide which tools and habits to prioritise among overwhelming 2025-2026 options.

Method

// When should you use the Jeff Su 3-Level AI System?

Use this skill when someone wants to build a sustainable AI workflow from scratch, upgrade their current ad-hoc AI usage into a systematic approach, or decide which tools and habits to prioritise among the overwhelming options available in 2025-2026.

// What do you need before starting the 3-Level AI System?

Current AI accessrequired
Which AI tools the user currently has access to (free or paid: ChatGPT, Claude, Gemini, or other)
Primary work typerequired
What the user primarily does — e.g. research, writing, coding, data analysis, video/media work, Google Workspace-heavy tasks
Technical comfort levelrequired
Whether the user is non-technical, somewhat comfortable with code, or a developer — determines which AI system tier is appropriate
Recurring work streams
2-3 recurring tasks or projects the user does repeatedly that would benefit from saved context
Existing tools/platforms
Tools where the user's context already lives — e.g. Gmail, Google Drive, Notion, Slack, email

// What core principles drive the Jeff Su AI System?

Go Deep on One, Carry Over to the Rest

Top models have converged in capability and copy each other's core features. Depth on one chatbot transfers directly to all others, so breadth-hopping is a trap — commit to one and build genuine fluency.

Paid Tier Priority

The gap between free and paid AI tiers is 'like night and day.' Always default to the most capable model available within your paid tier, because companies deliberately default you to the weakest (cheapest) model.

Context is King

With modern models, your prompt is no longer the biggest factor in output quality — the right context is. The right context will always beat the perfect prompt.

OC Framework (Outcome + Context)

The only prompting framework worth remembering: state a clear Outcome and supply the right Context. Models are now powerful enough to accurately infer role, format, and tone from context alone — you do not need to spell those out.

Projects as Permanent Homes

Projects (called Gemini Gems in Gemini) are permanent homes for recurring work streams. They store project instructions (rules and constraints), knowledge files (reference material and examples), and auto-updated memory — so you stop repeating yourself.

AI Systems Compound Over Time

An AI system connects multiple projects so it can cross-reference context across silos and update itself after feedback. Learnings compound: the more you use and correct it, the less instruction it needs going forward.

Reconcile Move

After AI produces a draft and you edit it, share your final version back and instruct the system to reconcile the two. The AI dissects every change, extracts rules, and applies them automatically to future outputs — accelerating style alignment over time.

// How do you apply the 3-Level AI System step by step?

1
Select your one chatbot from the Big Three
Choose from ChatGPT, Claude, or Gemini only. Apply three selection principles in order: (1) Paid Tier Priority — if your employer gives you paid Gemini but you have free ChatGPT, use Gemini; (2) Match to work type — ChatGPT for research/web search, Claude for writing/coding/design, Gemini for mixed media or heavy Google Workspace use; (3) Vibes — the one you enjoy most is the one you'll use most. Commit. Do not split attention across multiple tools at this stage.
2
Change your model default to the most powerful available
Navigate settings and manually select the highest-tier model you have access to. Companies default you to the weakest model. The most capable models break down requests, map steps, and catch nuances weaker models miss. Do this every session until it becomes habit.
3
Drop prompting frameworks and adopt the OC Framework exclusively
Stop memorising role/format/tone prompting templates. For every task, identify just two things: (1) Outcome — one clear sentence of what you want produced; (2) Context — the right input material that lets the AI infer everything else. Combine them in your prompt.
4
Source the right context using the three context-finding methods
Method 1 — Named Frameworks: explicitly name a recognised methodology (e.g. 'rewrite this using the pyramid principle'). If you don't know the right framework, ask the AI: 'What are the best frameworks for [goal]?' then pick one. Method 2 — Real Examples: paste 2-3 approved past outputs as examples rather than describing format/tone in prose; examples contain everything you forget to say. Method 3 — Connect Your Tools: point AI at the platforms where your context already lives (Drive, Notion, Slack, email) rather than manually downloading and re-uploading files.
5
Identify 2-3 recurring work streams and build a Project for each
A Project (or Gemini Gem) has three components you must populate: (1) Project Instructions — the always-applicable rules, goals, and constraints for this work stream; (2) Knowledge Files — source documents, real examples of good output, named frameworks. Use .md markdown files instead of PDFs wherever possible — they are easier for the AI to read and cheaper to process. Ask the AI to convert PDFs to markdown if needed. (3) Memory — leave this to the AI to update automatically as the work stream evolves. Do not try to pre-fill memory manually.
6
Apply the Reconcile Move to any writing project inside a Project
Workflow: AI produces draft → you edit it to match your voice/standards → paste your final version back → prompt: 'Reconcile my final version with your initial draft and propose rules to remember for next time.' The AI extracts the delta as explicit rules and applies them going forward. The more you do this, the fewer instructions you need.
7
Migrate Projects into an AI System once you have multiple active Projects
Choose your AI system tier based on technical comfort: (1) Non-technical / minimal setup — Gemini Spark (pre-connected to Gmail, Calendar, Drive; least control); (2) Non-technical / wants more control — Claude Cowork (requires some setup but no coding); (3) Comfortable with code — Claude Code or OpenAI Codex (fully customisable, maximum power). The model selector UI is a reliable signal: Gemini Spark has no model choice, Cowork has limited options, Codex exposes all options. An AI system enables two things Projects cannot: cross-project context synthesis and self-updating rules from feedback.

// What does the 3-Level AI System look like in practice?

A marketing manager who writes weekly status updates, runs campaign reports, and does competitive research — currently using free ChatGPT with no structured workflow

Step 1: Employer provides paid Gemini → switch to Gemini as primary. Step 2: Select most powerful Gemini model in settings. Step 3-4: For competitive research, paste competitor press releases and earnings calls as context (real examples method) + name a relevant framework ('analyse this using the Jobs-to-be-Done framework'). Step 5: Create three Projects — (a) Weekly Status Updates project with last 3 approved updates as knowledge files and a constraint in instructions like 'always under 200 words, bullet format'; (b) Campaign Reporting project with brand guidelines and past reports; (c) Competitive Research project with named frameworks and source documents. Step 6: After each draft edit, run the Reconcile Move to teach the system the manager's voice. Step 7: Once all three Projects are running, migrate to Gemini Spark so the system can cross-reference campaign spend data, calendar deadlines, and email threads simultaneously.

A non-technical operations professional who manually creates process documentation, tracks team tasks across Notion and email, and repeats the same onboarding prep every month

Step 1: Uses Claude on paid plan at work → commit to Claude. Step 2: Switch default to most capable Claude model. Step 3-4: For documentation tasks, paste two previously approved process docs as context (real examples method); prompt with clear outcome: 'Write a process doc for [new workflow] in the same format as these.' Step 5: Build a Recurring Onboarding Project — instructions contain constraints (audience = new hires, reading level, required sections); knowledge files contain past onboarding docs and team handbook in markdown. Step 6: After editing each new onboarding doc, run the Reconcile Move. Step 7: Once comfortable, set up Claude Cowork to connect the Onboarding Project with a separate Task Tracking Project — the system can then flag when onboarding timelines conflict with team capacity data from Notion.

// What mistakes should you avoid with the 3-Level AI System?

Spreading attention across multiple AI tools simultaneously instead of going deep on one — skills do not compound when you split focus.
Staying on the free tier when a paid tier is available through work or another route — the free-to-paid gap is too large to ignore.
Accepting the default model the platform gives you — companies default you to the weakest, cheapest model; always manually select the most capable one you have access to.
Over-investing in elaborate prompting frameworks and ignoring context quality — the right context will always beat the perfect prompt with today's models.
Describing what good output looks like in prose instead of pasting real examples of good output — examples contain everything you forget to explicitly say.
Uploading PDFs as knowledge files when markdown is available — PDFs are harder for the AI to read and more expensive to process; convert to .md first.
Treating Projects as the ceiling — Projects are siloed and cannot cross-reference each other; once you have multiple active Projects, the next step is an AI System.
Skipping the Reconcile Move after editing AI drafts — without reconciliation, the system never learns your preferences and you keep giving the same corrections forever.
Choosing an AI system tier beyond your current technical comfort — starting with Gemini Spark or Cowork is correct for non-technical users; jumping to Codex without code comfort creates friction that kills adoption.

// What are the key terms in the Jeff Su AI System?

The Big Three: The only three frontier chatbots worth choosing between: ChatGPT, Claude, and Gemini. All others are either non-competitive, derivative, or search tools rather than true frontier models.
OC Framework: The one prompting framework worth remembering — Outcome + Context. State a clear outcome (what you want produced) and supply the right context (material the AI needs to infer role, format, and tone). Replaces all multi-part prompting templates.
Context is King: The principle that, with modern AI models, the quality and relevance of context supplied to the model is a more decisive factor in output quality than prompt engineering or prompt length.
Named Frameworks: One of three context-finding methods: explicitly naming a recognised methodology (e.g. 'pyramid principle', 'Jobs-to-be-Done') in your prompt, because two words of a framework name carry more context than a paragraph of description.
Real Examples: The best type of context — pasting 2-3 actual approved past outputs so the AI learns format, tone, and unstated expectations directly from evidence rather than description.
Projects: A permanent home for a recurring work stream inside ChatGPT or Claude, containing three components: Project Instructions (always-applicable rules and constraints), Knowledge Files (reference documents and examples), and Memory (auto-updated by the AI to track milestones and key changes).
Gemini Gems: Gemini's equivalent of Projects — permanent, instruction-driven workspaces for recurring tasks within the Gemini ecosystem. Same structure, different name.
AI System: A setup that connects multiple Projects so it can (1) pull context across different Projects, spot cross-project patterns, and surface insights no single Project could find, and (2) update its own rules automatically when given feedback — so learnings compound over time.
Reconcile Move: A technique where you share your edited final draft back to the AI alongside its original output and ask it to reconcile the two, extract the rules behind every change you made, and apply those rules to future outputs — accelerating alignment with your voice and standards.
Gemini Spark: Google's most beginner-friendly AI system tier — pre-connected to Gmail, Google Calendar, and Google Drive with minimal setup required, but with the least user control over configuration.
Claude Cowork: An AI system option from Anthropic designed for non-technical users — provides more control than Gemini Spark and requires some setup, but no coding knowledge is needed.
Paid Tier Priority: The first model-selection principle: always prioritise whichever of the Big Three you have paid access to, because the capability gap between free and paid tiers is 'like night and day'.
.md markdown files: The preferred file format for knowledge files inside Projects — easier for AI to read and cheaper (in tokens) to process than PDFs. Any PDF can be converted to markdown by asking the AI to do it.

// FREQUENTLY ASKED QUESTIONS

What is the Jeff Su AI From Scratch 3-Level System?

It's a structured methodology for learning AI that moves through three levels: mastering one chatbot (ChatGPT, Claude, or Gemini) with the OC Framework, building Projects as permanent homes for recurring work, then scaling to an AI System that cross-references context and self-updates. It skips the irrelevant 80% of AI advice and focuses only on durable, practical skills.

What is the OC Framework in AI prompting?

The OC Framework stands for Outcome + Context — the only prompting framework worth remembering. You state a clear outcome (one sentence of what you want produced) and supply the right context (the material the AI needs). Modern models are powerful enough to infer role, format, and tone from context alone, so you no longer need multi-part prompting templates.

How do I choose between ChatGPT, Claude, and Gemini?

Apply three principles in order: prioritise whichever you have paid access to (the free-to-paid gap is night and day), match to your work type (ChatGPT for research, Claude for writing/coding, Gemini for Google Workspace and mixed media), then pick by vibes. Commit to one and build genuine fluency — depth on one chatbot transfers to all others.

How do I set up an AI Project for recurring work?

Identify 2-3 recurring work streams, then create a Project (or Gemini Gem) for each with three components: Project Instructions (always-applicable rules and constraints), Knowledge Files (source docs and 2-3 real examples of good output in markdown), and Memory (left for the AI to auto-update). This stops you repeating context every session.

How does this compare to learning AI through generic prompt engineering courses?

Generic courses over-invest in elaborate prompting templates that no longer matter with modern models. This system flips the priority: context beats prompts. Instead of memorising role/format/tone frameworks, you master one chatbot, supply real examples as context, and build Projects that compound over time — a durable workflow rather than a bag of prompt tricks.

When should I move from Projects to an AI System?

Move to an AI System once you have multiple active Projects that would benefit from cross-referencing each other. Projects are siloed and cannot share context; an AI System connects them so it can synthesise insights across silos and update its own rules from feedback. Choose your tier by technical comfort: Gemini Spark, Claude Cowork, or Claude Code/Codex.

What is the Reconcile Move and why does it matter?

The Reconcile Move is where you paste your edited final draft back to the AI alongside its original output and ask it to reconcile the two, extract the rules behind every change, and apply them going forward. It matters because without it, the system never learns your voice and you keep giving the same corrections forever.

Why should I use the most powerful model instead of the default?

Companies deliberately default you to the weakest, cheapest model to save costs. The most capable models break requests into steps, map nuances, and catch things weaker models miss — the difference is significant. Manually select the highest-tier model you have access to in settings every session until it becomes habit.

What results can I expect from using this system?

You'll stop repeating yourself, produce outputs that match your voice with fewer corrections over time, and build a workflow where learnings compound. Early on you supply lots of context and run Reconcile Moves frequently; over weeks the system needs less instruction. The endpoint is an AI System that cross-references your work and self-improves from feedback.

Why is context more important than the perfect prompt?

With modern AI models, prompt quality is no longer the biggest factor in output quality — the right context is. Models can infer role, format, and tone accurately when you supply real examples, named frameworks, or connected data sources. The right context will always beat the perfect prompt, so invest your effort there.

Should I use PDFs or markdown files for AI knowledge files?

Use markdown (.md) files wherever possible. Markdown is easier for the AI to read and cheaper in tokens to process than PDFs. If your source material is a PDF, ask the AI to convert it to markdown first before adding it as a knowledge file inside a Project.

// GET THIS SKILL — FREE