Most self-directed learning plans are a pile of links: a course, three articles, a YouTube playlist, maybe a book someone recommended. You work through some of it, feel like you learned something, and three months later cannot say precisely what you can now do that you could not do before.
A curriculum is a different object. It does not start with resources. It starts with a target capability, then works backward: what has to be true before you can attempt it, what tasks build it in practice, what evidence proves you can actually do it, and who checks that evidence before you trust it.
This article is a method for building that structure yourself, with AI doing the drafting and research legwork — and a warning about the two specific ways AI-generated curricula go wrong: invented credentials treated as real, and unsupervised practice tasks that are genuinely unsafe.
AI is fast at generating a plausible-looking sequence of “first learn X, then Y, then Z.” Plausible is not the same as correct. Every prerequisite claim, every named certification, and every practical task needs a verification step before you build your plan around it — because a fluent, confident wrong answer looks identical to a fluent, confident right one.
The four things a curriculum has to answer
A learning plan that only lists resources is missing at least three of these:
- Target capability — a specific thing you can do, stated as a task, not a topic.
- Prerequisites — what has to be true (skills, tools, access, foundational knowledge) before the target task is attemptable at all.
- Practice tasks with evidence — exercises that produce something checkable, not just hours spent consuming material.
- Review checkpoints — a point where a person, not the model that helped you plan, looks at the evidence and tells you whether you are actually ready to move on.
Miss the first and you drift between topics without a finish line. Miss the second and you attempt tasks you are not ready for, or waste time on foundations you already have. Miss the third and “I studied it” never becomes “I can do it.” Miss the fourth and you are the only judge of your own progress, which is the least reliable judge available.
Step 1: State the target capability as a task
Weak target: “Learn data analysis.” Usable target: “Given a messy spreadsheet export, clean it, compute the three metrics my team actually reports monthly, and present the result in a one-page summary a non-technical manager can act on.”
The difference is that the second version is checkable. You either produced that one-page summary from that kind of input, or you did not. The first version has no finish line, so no curriculum built on it can have one either.
Write your target capability as a sentence with a concrete input, a concrete action, and a concrete, checkable output. If you cannot picture what the finished thing looks like, the target is still a topic, not a capability.
Step 2: Map prerequisites — and verify them
Ask AI to draft a dependency chain, then treat the draft as a hypothesis, not an answer.
Target capability: [your one-sentence task from Step 1]
Draft a prerequisite chain: what has to be true — skills, concepts,
tools, access — before someone could reasonably attempt this task.
Order it from most foundational to most immediate. For each
prerequisite, mark your confidence as high, medium, or low, and name
the type of source that would confirm it (professional body,
textbook, official syllabus, practitioner).
Do not invent a certification, qualification, or standard body. If a
credential seems relevant but you are not certain it exists under
that name, say so explicitly instead of naming one.
Then verify anything you did not already know to be true. For domains with an actual credentialing structure — accounting, engineering, healthcare, teaching, most regulated trades — check the named body’s own site, not a summary of it. If the plan references a qualification you have not personally verified, search for the awarding body directly before you spend a single hour preparing for it.
Models sometimes generate a certification, exam, or professional designation that sounds exactly like a real one but does not exist under that name, or exists with different requirements than described. This is not a rare glitch; plausible-sounding institutional names are a known failure mode. Before you plan around any named credential, verify it independently: the awarding body’s own website, a national qualifications register, or for cross-border recognition of existing qualifications, a resource like the EU’s ENIC-NARIC network. If you cannot find independent confirmation the credential exists, do not build a study plan around it.
For prerequisite chains outside formal credentials — “you need to understand X before Y makes sense” — the verification bar is lower but not zero. Cross-check against an actual syllabus, textbook table of contents, or a practitioner’s own account of how they learned it, rather than accepting the model’s ordering as ground truth. Sequencing is exactly the kind of claim that sounds authoritative and can still be wrong.
Step 3: Design practice tasks that produce evidence
A practice task is not “read chapter three.” It is an action that leaves behind an artifact you or someone else can check: a piece of code that runs, a document that follows a spec, a repaired object that works, a conversation you can point to. Time spent consuming material is not evidence of capability; the artifact is.
For this prerequisite or sub-skill: [name it]
Design one practice task that:
- takes under 90 minutes;
- produces a concrete artifact (not notes or a summary);
- has a clear, checkable definition of "done";
- states what a reviewer with no context would need to see to judge it.
Flag if this sub-skill cannot be safely or usefully practiced without
in-person supervision, and say why.
Push for a checkable “definition of done” every time. “Practice interviewing techniques” produces nothing to check. “Record a 10-minute mock interview using this question set, and mark the three moments where you interrupted or filled silence” produces an artifact — the recording plus your own annotation — that a reviewer can actually look at.
Some practical tasks are genuinely unsafe to attempt from an AI-generated description alone: anything involving electrical work, gas, structural changes, medical or first-aid procedures, operating vehicles or machinery, chemical handling, or working at height. A model has no way to see your specific setup, verify your tools, or confirm local safety codes, and a confidently written five-step guide reads the same whether or not it is missing a critical safety step. For any task in a safety-relevant domain, treat the AI-drafted version as a starting outline only, and get supervised, in-person instruction before you practice it unsupervised — the honest answer for “how do I practice this safely” is very often “with a qualified person the first several times,” not “alone, from a generated guide.”
Step 4: Set review checkpoints
Every three to five practice tasks, or at the end of a major sub-skill, build in a checkpoint where a human other than you looks at the accumulated evidence. Options in rough order of rigor:
- a knowledgeable friend or colleague reviews two or three artifacts and tells you honestly whether they would trust the work;
- a practitioner community (forum, local meetup, professional association) critiques a specific piece of work;
- a paid mentor or tutor reviews a portfolio of artifacts at a scheduled interval;
- for regulated or credentialed skills, an actual assessment through the recognized body.
The checkpoint’s job is to catch what self-assessment cannot: blind spots, bad habits practiced into muscle memory, and the gap between “this felt right to me” and “this is actually correct.” AI can help you prepare for a checkpoint — organizing the artifacts, drafting questions to ask the reviewer — but it cannot be the checkpoint. It was not there when the artifact was produced, has no way to independently verify claims about it, and, notably, tends toward agreeable feedback rather than the pushback a real checkpoint is for.
A worked example
Target: “Given a customer complaint email, draft a response that resolves the issue, follows company tone guidelines, and needs no more than one round of manager edits.”
- Prerequisites (verified): understanding of the specific product/policy area; company tone guide (an actual internal document, not a model’s guess at one); basic de-escalation language patterns (cross-checked against the company’s own training material, not invented).
- Practice task 1: rewrite three real (anonymized) past complaint emails against the actual tone guide; artifact is the three rewrites plus a one-line note on what rule each one follows.
- Practice task 2: role-play a difficult complaint with AI as the escalating customer; artifact is the transcript plus your own note on where you nearly over-promised.
- Checkpoint: a manager reviews five of your drafts against real incoming complaints for two weeks and tracks the edit rate.
Nothing here needed an invented certification, and nothing needed unsupervised practice on anything unsafe — the whole plan is built from real internal material and a human reviewer who already exists in the workplace.
Common failure modes
- A reading list dressed up as a curriculum. If every item is “read/watch X,” there is no practice task and no evidence, only exposure.
- An unverified prerequisite chain treated as authoritative because it was generated confidently and in order.
- A named certification nobody can find when searched independently.
- Practice tasks with no definition of done, which means no way to tell progress from motion.
- No human checkpoint at all, so the only judge of “am I ready” is the same model that drafted the plan and has no way to check the artifacts against reality.
The honest limit
AI is genuinely useful for drafting a first-pass curriculum structure quickly, generating varied practice tasks, and helping you prepare for a review conversation. It cannot validate that a prerequisite chain is correct for your specific field, confirm that a credential exists or means what it claims to mean, or replace a competent person judging your actual work. Use the skill curriculum schema to fill in a target capability, its verified prerequisites, three to five practice tasks with evidence, and at least one scheduled human checkpoint before you invest real hours — and treat anything the model could not point you to an independent source for as still unverified, not as a fact.
Two adjacent workflows are worth combining with this one: the 30-day AI learning plan for pacing the actual study sessions, and deep research mode for the verification step itself, since finding and checking real sources is exactly the kind of multi-step lookup it is built for.
