TL;DR
- Paste the prompt below into ChatGPT (GPT-5.5), Claude (Sonnet 4.6), or Gemini 3.1 Pro to turn a vague goal into a 12-week path with one shippable milestone per week.
- The single best change to any AI-generated plan: cap each week at 1 video + 1 chapter + 1 exercise, and make every milestone something a friend could verify in 60 seconds.
- Use a chat model to design the path, then use a notebook tool (Google NotebookLM, free, up to 50 sources as of June 2026) to hold the actual study material, and Anki for spaced repetition. No single tool does all three well.
- Build the exit checkpoints (weeks 4, 8, 12) before you start. They are the only thing that catches a wrong path before it wastes three months.
The task
It’s the first Sunday of the month and you’ve decided, for real this time, to learn data science (or design, or Spanish, or indie hacking). Three browser tabs are already open: a Coursera specialization, a Reddit “best resources” thread, and a 50-item Notion template called “Roadmap 2026.” You don’t need more inputs. You need a 12-week path with weekly milestones you can actually finish on a Saturday morning, three honest checkpoints to catch yourself before week 12, and a “things to skip” list so you don’t spend the first week configuring a perfect dev environment instead of writing code.
Where AI helps, and where it does not
AI is excellent at assembling a coherent sequence from common resources, converting vague goals into shippable weekly deliverables, and naming the things beginners over-invest in (Vim configs, perfect Anki decks, choosing a framework before writing one line). It can also predict where you’ll plateau (usually weeks 5-7) and bake in a checkpoint. What AI cannot do: pick the resource that matches your learning style. Two people with the same goal need different first-week material — one needs a single video walkthrough, the other needs a written reference plus exercises. Try 2 resources for the first concept and pick.
The named failure mode: the curriculum dump. AI lists five courses for week 1, three books for week 2, and a 200-hour course for the back half. Beginners read the list, feel overwhelmed, ship nothing. Force the prompt to cap resources at one video + one chapter + one practice exercise per week.
Which model for which step
Any of the three big chat models writes a solid 12-week path. The differences are at the edges, current as of June 2026:
| Step | Tool | Why | On-site guide |
|---|---|---|---|
| Draft the 12-week path | ChatGPT (GPT-5.5), Claude (Sonnet 4.6), or Gemini 3.1 Pro | All three handle structured planning well; free tiers are enough for one path | ChatGPT vs Claude vs Gemini |
| Plan from a long syllabus or textbook PDF | Gemini 3.1 Pro or Claude Sonnet 4.6 (both 1M-token context as of June 2026) | Paste an entire syllabus or several chapters without truncation | ChatGPT vs Claude vs Gemini |
| Hold and quiz the actual study material | Google NotebookLM (free, up to 50 sources) | Grounds answers in your uploaded notes; generates a study guide and audio overview | NotebookLM getting started |
| Spaced-repetition review | Anki + AI-generated cards | NotebookLM and chat models do not track what you know or schedule reviews; Anki does | Make flashcards with AI |
Free ChatGPT, Claude, and Gemini tiers all comfortably generate a single plan. NotebookLM is free with a Google account and caps free notebooks at 50 sources each; that is far more than one learning path needs. Keep planning (chat model) separate from material storage (NotebookLM) separate from drilling (Anki). One tool that claims to do all three usually does the third one badly.
What to feed the AI
- The field, plus the specific outcome you want at week 12 — concrete and verifiable, not aspirational (“ship a Streamlit dashboard reading from one CSV”, not “understand SQL”)
- Realistic hours per week — actual, not aspirational (account for the gym, the kids, the Sunday hangover)
- Existing background that overlaps with this field (programming, math, a language, design intuition)
- Your accountability constraint — solo, paid coach, study buddy, public Twitter
- Resources you already own (books, course logins) so AI doesn’t recommend redundant ones
- Learning style preference if you know it — video, written, hands-on, paired
- The single artifact you would point at and say “I learned this” at week 12
- Hard constraints — exam date, job interview, conference talk, public commitment
Copy-ready prompt
Design a 12-week self-study path for me.
Field + concrete week-12 outcome: {field, plus a deliverable I can show to a friend in 60 seconds}
Realistic hours per week: {n hours, no aspirational numbers}
Existing background: {what I already know that overlaps}
Resources I already own: {list}
Accountability constraint: {solo / paid coach / study buddy / public}
Hard date constraint: {if any}
Return:
1) Week-by-week milestone — each must be a concrete deliverable I can show to a friend in 60 seconds. Not "learn X", but "ship Y" or "explain Z out loud."
2) Resource recommendation per week — cap at 1 video + 1 book chapter + 1 practice exercise. No more. If two resources teach the same thing, pick the one that practices the milestone.
3) Exit checkpoints at weeks 4, 8, and 12 — phrased as "if I cannot do X without looking it up, pause and re-plan." Make them specific and verifiable.
4) "Things to skip" list — the 5 things beginners in this field over-invest in (tooling, framework debates, configs, perfect notes, taxonomy memorization).
5) Plateau warning — name the week where most learners stall and the one concrete unblock for that plateau.
6) Community / accountability suggestion that matches my constraint above.
Rules: no week may list more than 3 resources. Every milestone must be shippable, not memorizable. If two weeks share a milestone, merge them and add a stretch week.Shorter variant — 4-week sprint
4-week sprint plan for {field}.
Hours/week: {n}. Background: {bg}. Sprint goal: {one shippable artifact at week 4}.
Each week: 1 milestone (shippable), 1 resource, 1 hour budget. Plus a "skip this" list and a week-2 plateau unblock.Sample output
A useful milestone (data science): “Week 3 — ship a one-page Streamlit app that reads from a CSV and renders one chart. Push to GitHub. DM the link to one friend.” This beats “Learn pandas” because the first is verifiable in 60 seconds; the second is not.
A useful checkpoint: “Week 4 checkpoint — without looking it up, write a 5-line pandas snippet that filters a CSV by one column and groups by another. If you cannot, the prior 3 weeks were too video-heavy. Switch the next 4 weeks to exercises-first.”
A useful “skip” list (programming-adjacent): “Skip: configuring a personal Vim/VSCode theme; picking the ‘best’ Python version manager (use whatever your tutorial uses); reading a second linear algebra textbook before writing code; learning Docker in week 1; deciding on a portfolio site before you have 1 project.”
A useful plateau warning: “Most learners stall in weeks 5-7, when concepts stop being intuitive and the practice exercises feel longer. Unblock: drop new material for one week and re-ship two prior weeks’ milestones from scratch without looking. The plateau is not knowledge; it is integration.” This re-shipping move is just retrieval practice and spaced repetition, two of the few study techniques with strong evidence behind them (Dunlosky et al., Psychological Science in the Public Interest).
How to refine
- Make milestones shippable, not memorizable: “Every milestone must be something I can show to a friend in 60 seconds. Replace anything that starts with ‘understand’ or ‘learn’.”
- Cut redundant resources: “Each week, if two resources teach the same concept, remove the one that does not include practice. The remaining resource should produce the milestone.”
- Make checkpoints fail-able: “Phrase each checkpoint as a task I either can or cannot do without looking it up. Vague checkpoints get rationalized into passes.”
- Lengthen the ‘skip’ list: “Add 3 more items to the skip list. Be specific — name tools, books, or YouTube rabbit holes, not categories.”
- Plan the plateau: “Where will I stall, and what’s the one tactical unblock? Bake it into week 5 as a ‘review week’ if needed.”
Common mistakes
- Listing 5 courses for week 1 — paralysis, no shipping, no signal of progress by Sunday night
- No exit checkpoints — you don’t notice the path is wrong until week 12 when the outcome doesn’t ship
- Skipping the “things to skip” list — beginners over-invest in setup, tooling, and taxonomy in the first month
- Aspirational hour counts — planning for 10 hours/week when the honest number is 4 sets the plan up to fail by week 3
- Milestones that are about consumption (“watch chapter 5”) instead of production (“ship a thing using chapter 5”) — consumption is invisible
- Public-learning when private-learning is the actual mode — performing kills the practice loop
- No catch-up slack — weeks 4 / 8 / 12 should be lighter so missed weeks have somewhere to go
- Picking a final outcome that’s too vague (“get good at SQL”) — vague outcomes mean vague evidence at week 12
FAQ
- What if I miss a week?: Use the catch-up slack already built in at weeks 4, 8, 12. Do not redistribute the missed week across all remaining weeks — that compounds the slip into burnout.
- Should I share my plan publicly?: Yes if public commitment increases the chance you ship the week-12 outcome. Skip if public learning makes you perform-and-share instead of practice-and-stumble. Practice loops require allowed failure.
- How specific should the week-12 outcome be?: Specific enough that a stranger could verify it in 60 seconds. “Ship a Streamlit dashboard reading one CSV with one chart, public link.” If you can’t verify it in 60 seconds, the goal is too soft.
- Should I do two parallel paths (e.g., language + design)?: Usually no. Two parallel 5-hour paths almost always become one 7-hour path with guilt. Pick one for 12 weeks, then stack.
- AI keeps recommending paid courses I cannot afford. How do I bias it toward free?: Add to the prompt: “Free / library-borrowable resources only. If the best resource is paid, suggest the closest free equivalent and note the gap.” Most fields have an 80%-as-good free path.
- Which AI should I use, and do I need to pay?: For a single 12-week plan, the free tiers of ChatGPT (GPT-5.5), Claude (Sonnet 4.6), or Gemini 3.1 Pro are all enough, as of June 2026. If you want to paste a whole syllabus or several textbook chapters in one go, Gemini 3.1 Pro and Claude Sonnet 4.6 carry a 1M-token context so nothing gets truncated. Use Google NotebookLM (free, up to 50 sources) to store the actual material and quiz yourself against it, and Anki for the spaced-repetition reviews the chat models cannot schedule for you.
- Can AI keep me on track week to week, not just plan the path?: Partly. A chat model will re-plan when you paste your honest progress (“I shipped weeks 1 and 2, skipped 3”), and NotebookLM will quiz you on your own notes, but neither nudges you on a schedule. The accountability constraint you set in the prompt (study buddy, public commitment, paid coach) does more for follow-through than any tool.