How to Learn AI in 2026: A 90-Day Self-Study Plan
A week-by-week structured plan. Not another list of courses to watch passively, a build-first curriculum with clear milestones, free resources, and three portfolio projects you ship before day 90.
Why this plan is different
Most "learn AI" content gives you a list of courses. This gives you a plan with milestones, projects, and a decision framework for what to study next. The difference matters: passive course watching without projects produces learners who cannot actually build. This plan is build-first, you ship working code every two weeks, starting in Week 3.
The plan is calibrated for 2026 specifically. Learning AI in 2026 means learning LLM application development first, using Claude, OpenAI, and Gemini APIs to build things, rather than the traditional path of studying statistics and training neural networks from scratch. That traditional path is valuable for research careers, but the LLM API path is where the jobs and product opportunities are today.
90 days is realistic. Not fast, there are no shortcuts to actually understanding what you're building, but achievable with consistent effort. The single biggest predictor of success is not prior technical experience but weekly consistency: 10 hours per week, every week, for 13 weeks.
What you need before Day 1
Required
- A computer with internet access
- 10 hours per week of consistent study time
- Willingness to look things up when stuck (Google, Claude, Stack Overflow)
- A free Claude account or OpenAI account for API access
Helpful but not required
- Any prior programming experience (speeds up Phase 1 by 1-2 weeks)
- A domain you want to apply AI to (healthcare, writing, finance, etc.)
- High school math (algebra, basic statistics)
No Python? Add one week before Phase 1. Work through the first two modules of Python for Everybody (Coursera, free to audit) or Codecademy's Learn Python 3 before starting the plan. 10 hours of Python fundamentals is enough to follow the Phase 1 coding exercises.
The 90-Day Plan
Five phases, each building on the last. Phases 1-4 each run 2-3 weeks. Phase 5 is the final week, agent patterns plus portfolio polish. Every phase ends with a concrete project you ship.
Foundations, What AI Actually Is
Goal: Understand what AI can and cannot do, use at least one AI tool confidently, write your first working Python script
Free Resources
Milestones
- Explain what LLMs are and how they work at a conceptual level
- Create a Claude account and use it productively for a real task
- Write a Python script that reads a file and prints its contents
- Call the Claude or OpenAI API using Python and get a response
Phase Project (ship this before moving on)
Build a simple Python script that calls an LLM API and summarizes a news article you paste in.
Prompt Engineering, Getting Reliable AI Outputs
Goal: Write prompts that produce consistent, high-quality outputs for a real use case you care about
Free Resources
Milestones
- Understand and apply chain-of-thought prompting
- Write a few-shot prompt that reliably formats output as JSON
- Design a system prompt for a specific assistant persona
- Evaluate 3 different prompt variations for the same task and explain why one wins
Phase Project (ship this before moving on)
Build a prompt library of 10 tested prompts for a domain you know well (your job, your hobby, a subject you study). Test each prompt 5 times and refine until outputs are consistent.
Building with LLMs, Your First AI Application
Goal: Ship a working AI application that someone else can use
Free Resources
Milestones
- Build a chatbot with conversation history using the Claude or OpenAI API
- Add a system prompt and streaming responses to your chatbot
- Deploy the chatbot to a URL others can access (use Vercel or Streamlit)
- Collect 5 pieces of user feedback on the deployed application
Phase Project (ship this before moving on)
Build and deploy a domain-specific chatbot. Examples: a customer support bot for a fictional product, a writing assistant for a specific style, a study helper for an exam you're preparing for. Must be live at a URL.
RAG, Making AI Know Your Data
Goal: Build a system that can answer questions about any set of documents you provide
Free Resources
Milestones
- Explain what embeddings are and why they enable semantic search
- Chunk and embed a set of documents into a vector store (Chroma or Pinecone)
- Query the vector store and inject retrieved chunks into an LLM prompt
- Evaluate whether your RAG system answers questions your chatbot (Phase 3) could not
Phase Project (ship this before moving on)
Build a RAG chatbot over 10+ documents you own (course notes, saved articles, a textbook PDF). Must correctly answer at least 8 of 10 test questions you write before considering it done.
Agentic AI + Portfolio Polish
Goal: Understand agentic patterns, build one agent-powered demo, and publish your portfolio
Free Resources
Milestones
- Build a simple agent that uses at least 2 tools (e.g., web search + calculator)
- Understand the ReAct pattern (reason + act loops)
- Publish all 3 projects to GitHub with clear READMEs and live demo links
- Write a 1-page reflection on what you built, what broke, and what you would do differently
Phase Project (ship this before moving on)
Build an agent that can research a topic on the web and produce a structured report. Use Claude as the reasoning backbone with a web search tool and a document writer tool. Share the GitHub repo publicly.
A realistic weekly schedule
Ten hours per week sounds like a lot but breaks into manageable chunks. Here is what a typical week looks like across Phase 1-4:
| Day | Activity | Hours |
|---|---|---|
| Monday | Watch / read the week's learning resource | 2h |
| Tuesday | Implement the week's concept in code, follow examples, then try without looking | 2h |
| Wednesday | Work on the phase project, make measurable progress | 2h |
| Thursday | Debug what broke on Wednesday, read docs, ask Claude for help | 1h |
| Saturday | Polish and ship the week's work, review milestones checklist | 2h |
| Sunday | Optional: read a blog post or paper in the field. Prep for next week. | 1h |
| Total | 10h |
What to do when you're stuck
Getting stuck is part of learning. The difference between learners who finish and those who quit is usually not intelligence, it's the ability to systematically get unstuck. Here is the order to try when you hit a wall:
Read the error message carefully
AI error messages are usually informative. Read the full traceback, not just the last line. The error tells you what went wrong and often where.
Ask Claude or ChatGPT to explain the error
Paste the full error + the relevant code and ask the AI to explain what's wrong and how to fix it. This is the fastest debugging tool available to you.
Search for the exact error message in quotes
Someone else has hit this exact error. Searching the error message in quotes finds Stack Overflow answers, GitHub issues, and blog posts with fixes.
Check the documentation for the library you're using
Library docs often have a 'common errors' or 'troubleshooting' section. The Claude API docs and LangChain docs are particularly well-organized.
Step back and rebuild the minimum working example
If nothing works, strip the code to the simplest possible version that demonstrates the problem. Often the error becomes obvious when you remove distractions.
All free resources, organized by phase
| Resource | Phase | Format | Hours | Certificate |
|---|---|---|---|---|
| Anthropic Academy (full curriculum) | All phases | Video + text | 20-40h | Free |
| AI for Everyone, Andrew Ng | Phase 1 | Video | 6h | Paid |
| Google Generative AI Learning Path | Phase 1 | Video + quiz | 8-12h | Free badge |
| DeepLearning.AI Prompt Engineering | Phase 2 | Video + notebook | 2-3h | Free |
| Microsoft GenAI for Beginners | Phase 3 | Text + code | 15-20h | None |
| DeepLearning.AI: LangChain for LLM Dev | Phase 4 | Video + notebook | 3-4h | Free |
| DeepLearning.AI: AI Agents in LangGraph | Phase 5 | Video + notebook | 3-4h | Free |
| Fast.ai Practical Deep Learning | After Day 90 | Video + Jupyter | 25-30h | None |
After Day 90: what's next?
By Day 90, you have 3 deployed projects, completion certificates from Anthropic Academy and likely one other provider, and a foundational understanding of LLM APIs, prompt engineering, RAG, and agentic patterns. Here are three paths forward depending on your goal:
Research path
Go deeper: Fast.ai Practical Deep Learning, then Stanford CS231n (computer vision) or CS224n (NLP). Add mathematics with 3Blue1Brown's linear algebra series. Goal: understand model internals.
Engineering path
Build production-grade AI systems: add LangGraph for agentic workflows, Langsmith for evaluation, and a vector database at scale. Contribute to an open-source AI project on GitHub.
Product path
Apply your skills to a specific domain. Build a real product with real users. Add a Coursera Google AI Essentials certificate for non-technical roles. Start shipping for customers.
Related AI Courses Guides
Frequently asked questions
Ready to start?
The best time to start your 90-day AI plan is today. Open Anthropic Academy and complete the first module, it takes under 2 hours and gives you a working mental model of how Claude works.