AI-Assisted Learning Workflow

An AI-assisted learning workflow is a way of studying in which a model does part of the work and the learner keeps the part that becomes memory. A model can take on a large share of the work in a study session without the learner losing what the session is for. The split between the model’s work and the learner’s work is predictable. Work on the form of the material can go to the model: finding it, converting it to another format, and writing questions after the learner has already tried to answer from memory. Work that decides structure stays with the learner: what is important, what belongs with what, and how the pieces connect. That structural work is what becomes the learner’s memory of the subject.

Two checks during a session show which side the work landed on: whether the learner can start drawing how the pieces connect, and whether the learner can invent a hard question about them. One check at the end of the session decides it: whether the model sped up the learning or did the learning.

Core takeaways

• A useful goal names the final artifact. “Learn AI” is a weak goal. “Build a simple agent” names a thing that can exist on a screen when the work is done.
• The learner keeps relevance, organisation, schema formation, and final judgment. The model gets resource discovery, format conversion, section extraction, quiz generation, clarification dialogue, note cleanup, and artifact scaffolding.
• The learner builds a first model of the material before asking the model to organise, rank, or synthesise anything. The model may critique a map the learner drew. The model does not draw the map first.
• Coverage per minute and the feeling of understanding are recognition. Neither predicts recall or application. The map test and the hardest-question test show whether the structure is in the learner’s head.
• The end-of-session check is one question: did AI accelerate the learning, or did it perform the learning. If the learner cannot explain, modify, defend, or apply the artifact, the loop is not finished.
• Importance, grouping, and relationships stay with the learner because those three operations are the higher-order capability, and their value is rising.

Name the artifact first

A useful goal names the final artifact. A vague goal leads to vague learning. The brain discards information it cannot connect to a purpose, so the goal decides which information is relevant and what kind of finished thing will show the learning happened. Weak goals stay as abstractions: learn AI, understand finance, get better at coding, study Japanese. Stronger goals name a thing that can exist on a screen or in a room when the work is done: build a simple agent, produce a personal finance strategy, pass a specific exam, create a dashboard, hold a 10-minute spoken conversation, compile a wiki page from a source cluster. Without that named thing, research goes on without end and comprehension has no target.

The same reason sets the order of the loop. Problem-First Learning is the vault page that covers this rule: the problem a tool exists to solve comes before the tool. A tool becomes meaningful when it solves a problem. Teaching the tool first, as a list of parts, leaves the learner with no reason to keep any of it. The example behind the rule is an engine on a workbench. A course on screwdrivers and a course on wrenches leave both tools as objects with no use attached. When the learner is handed the engine and told to take it apart, the screwdriver becomes necessary at a specific moment, and its use is obvious. For this wiki, source processing starts with the page or capability wanted; the source pile exists to serve that target.

The five names and the seven steps

AI makes a good learning framework more productive, and the framework is still needed. The coarse frame has five names: Goal, Research, Priming, Comprehension, Implementation. Most of the clock goes to comprehension and implementation. The working model has seven steps, because two of them carry the division of labour that the five names leave out.

1. Name the artifact. Decide what should exist when the learning is done. Decide whether producing it leaves a harder capability than the one already held, or only a faster version of the same work.
2. Find the resources. AI searches communities, examples, courses, docs, and prerequisites. Ranking their importance stays with the learner.
3. Prime the field. A study guide, a pre-quiz, key terms, and a rough structure make the material less foreign. Priming is a short pass before studying properly. It produces a rough map. It is not the learning.
4. Build a first human model. The learner decides what matters, what groups together, and what seems unclear, before any request to organise, rank, or synthesise.
5. Use AI for format and feedback. Convert the medium, extract named sections, explain a gap the learner has already attempted, quiz, critique a map the learner drew, and make notes legible after the grouping exists.
6. Implement alongside learning. The report, wiki page, app, dashboard, deck, or strategy starts while the material is still being worked, not after.
7. Audit the division of labour. Did AI accelerate the learning, or did it perform the learning.

What the model can do and what the learner keeps

The learner keeps relevance, organisation, schema formation, and final judgment. Schema formation is the internal structure of a subject: what the pieces are, which ones matter, how they group, and how they connect. That work stays with the learner. AI is strongest at seven jobs around that work: resource discovery, format conversion, section extraction, quiz generation, clarification dialogue, note cleanup, and artifact scaffolding. The biggest gains are in two places: research that finds better resources sooner, and comprehension support that converts the format without removing active processing.

Research: widen the search, keep the ranking

Research with AI widens and filters the search space. The learner still chooses what matters. AI is useful for finding the missing pieces: how practitioners learned the topic, which courses or resources match the named artifact, common prerequisites, examples close to the desired output, and searches deeper than a normal query. The risk at this stage is asking the machine to decide importance before the learner has any model of the subject. The Right vs Wrong Way to Work With AI is the companion page that states the same rule as a prohibition: never ask for the answer; ask for the information that helps the learner figure the answer out. The order that page gives is feeling first, then thought, then question. Seeding a search with keywords is legitimate. Asking for an importance ranking is not. The material on that page dates to late 2024 and early 2025.

Priming: a short pass before the real study

Priming makes the material less foreign before depth work begins. A generated study guide, a pre-quiz, a list of key terms, simple definitions, a structural summary of a course or document, and a look at starter code or example artifacts all belong here. Definitions stay in the simplest terms that are still correct; a technically perfect complicated definition is worse. Extracted key terms are raw material for the learner’s own grouping. The same restriction applies: do not ask the machine to rank them. A structural summary is a hypothesis to confirm against the source. It is not a map to memorise.

Taking a quiz before learning is useful even when the score is poor. The unanswered question prepares the brain to notice the answer when it appears. The benefit is clearest for material the pretest touched. Generalisation to untested material is weaker, so a pre-quiz does not replace previewing the whole scope.

Prestudy is the priming step worked out in full. Priming is done when the topic has a big-picture structure rather than a term list, the whole subtopic is covered, no details are memorised yet, the major ideas have visible relationships, the learner has hypotheses about connections, and the main study feels like the next step. Perfect notes, flashcards, polished maps, or false confidence are not signs that priming is done.

Comprehension: where most of the time goes

Comprehension is where most learning time is spent, and where AI can help most if the boundary is clear. The model can convert text into single-speaker audio. It can convert video or audio into text. It can extract only the section relevant to the goal; the learner then checks the extract against the original’s table of contents. It can offer alternate examples of the thing itself. First-contact analogies to another domain stay off, because the learner has no way to evaluate them. The model can answer a specific question after a first attempt, once the gap can be named. Clarification dialogue has the same boundary: useful after that first attempt; a substitute for first contact when it comes before.

Diagrams and tables are grouping decisions in visible form. The learner draws the map, badly if needed. The machine may then critique it or answer a named question about it. The model does not generate the map before the learner has tried. The model may critique the map after the learner has tried.

Note cleanup after the session may make the page legible. The grouping is still the learner’s job. Notes already organised in the learner’s head can be cleaned at no cost. Notes that are not organised in the learner’s head hide that fact once they look tidy. Tidy notes on the page do not show that the material is organised in the head.

Format conversion is not automatically a shortcut. It becomes one when the machine decides importance, grouping, and relationships before the learner has tried. Comprehension itself runs in three layers: logic, concept, and detail. Sources present those three layers mixed together. That claim belongs to Layers of Learning.

Implementation: build the artifact while studying

Implementation is part of learning. It is not a step that comes after. The final artifact shows whether the knowledge can act. The goal shapes what gets noticed, skipped, and tried as the material is studied, so implementation is not fully separate from comprehension. By the time comprehension is done, the artifact should already be partly formed. AI can speed up that artifact.

Goal type	Useful AI support
Essay or report	outline, draft structure, critique
Application or code	scaffold, debugging, implementation help
Dashboard	data cleanup, visualization, interface generation
Slide deck	structure and first draft
Wiki page	source organization, outline, draft cleanup

If the learner cannot explain, modify, defend, or apply the artifact, the loop is not finished.

How to check whether the model did the learning

The end-of-session check is one question: did AI accelerate the learning, or did it perform the learning. Asked only after the artifact exists, the honest answer is often that the two look the same. So three checks, each about five seconds long, run while the session is still open.

The first check is on fluency. Coverage per minute and the feeling “I understand this” are fluency signals. Both are recognition. Neither predicts recall or application. A deck of cards shows the split: every card is recognised on sight, and the eight of hearts cannot be drawn from memory. Anything that makes the session feel like it is moving faster, such as a clean summary appearing instantly or faster playback, feels like learning. It is not evidence of learning.

The map test is the second check. The learner tries to start drawing how the pieces connect. If the start will not come, the structure is not in the head, however good the notes look. If the learner wants to redraw the map halfway through, the process is working.

The third check is to write the hardest applied question the material can bear. A learner who can answer found questions and cannot invent a hard one has a conceptual gap.

Those three checks make a clean artifact with weak encoding detectable during the session, rather than as a verdict afterwards. They are also the local form of The Shortcut Problem: a task demands higher-order work, the load is felt, an easier path completes a visible study behaviour, and the result looks correct while the knowledge does not improve. The test is whether relevance, organisation, and judgment stayed with the learner. It is not a list of rules to recite. Don’t Outsource the Learning is the evidence page for the same split: engineers who asked conceptual questions scored higher on later comprehension than engineers who pasted generated code from the same model. The tool was the same. The way each group used it was different. Shipping the artifact and learning the material are two separate measures.

Why format conversion is still allowed

Format conversion stays legitimate. The reason that used to go with it does not hold. Matching instruction to a declared learning style produces no learning advantage, and attention drifts easily during audio delivery, across the whole population. Learning Styles Myth and Multimodal Learning is the vault’s position. The reasons that still hold are different: a commute or a walk becomes usable time, a fixed medium can be met again in a second form, and the cost of starting drops. What goes on in the brain is what matters, whichever medium delivers the material. Faster playback is a coverage tool for the priming pass. It does not speed up comprehension. Watching faster does not make the learning faster; fluency tracks confidence, not retention.

The skip test

The skip test is one question: do I know enough about this to make it simpler. If the material cannot be simplified while staying accurate, the foundation is missing. The move is to drop a level, learn that level fully, and come back. The skipped items get written down so they are actually revisited. Advanced material studied without its foundation does not stay learned, however many times it is studied.

The machine can be wrong

Priming output is a hypothesis to confirm against the source. It is not a map to memorise. Extracted sections get checked against the original’s table of contents. Anything the model cites gets opened. Models lean too heavily on sources with many citations and invent citations for niche queries. Both faults land on the research step.

Apply on the first pass

Application starts on the first pass. The working cadence is a thought about applying the new information within five to ten minutes. That figure is a coaching default. It has not been measured. Information with nowhere to fit gets dropped. This is the same purpose filter that the goal step set up, now running continuously.

Mixing subjects across a day

Mixing unrelated subjects across a day, such as an hour of one language, an hour of finance, and two hours of agents, can help motivation and sustain attention. That is variety. The interleaving effect is a different thing: related, easily confused items mixed inside one domain, so the learner has to select the approach rather than only execute it. Spaced Interleaved Retrieval and Interleaving for Complex Problem Solving cover the real effect.

When the workflow fits and when it does not

The learner’s energy matters more than the hours available. A two-hour learning block after a draining day may produce less than a shorter block when the brain is fresh. Comprehension and implementation belong in high-energy windows when those windows exist.

The workflow fits when learning has a clear output, when the source pile is large and needs filtering, when AI help is wanted without handing over schema formation, when the friction of a medium is slowing comprehension, when notes are messy and the material has already been processed, and when a raw transcript needs turning into a first structured brief. It is not permission to let AI build the schema from scratch. The loop works only while relevance, organisation, and judgment stay with the learner.

The case against the workflow is three situations. First, no artifact is in mind; how the workflow adapts to exploratory learning is still an open question. Second, the first-model step was skipped, so encoding has not been attempted. Third, the workflow is being treated as permission to hand the work over. The price of the workflow is the first-model step and the mid-session checks, time that the default way of using an AI product does not spend. Three signals say to quit: a session that produces a clean artifact the learner cannot explain, modify, defend, or apply; a priming map accepted without opening the source; and two sessions in a row where the machine ranked, grouped, or related before a first attempt. The expectation that can be checked is the four verbs, explain, modify, defend, and apply, plus the map test.

The success metric that used to go unstated was hours saved. The reason to keep importance, grouping, and relationships with the learner is that those three operations are the higher-order capability, and they are the part of the loop whose value is going up. The goal step therefore carries a second question: does producing the artifact leave a harder capability than the one already held, or only a faster version of the same work.

How Top Performers Learn is the systems frame around that question: top learners design a personal system rather than hunting for one best technique.

The fetching, the converting, and the tidying can stay with the machine. Importance, grouping, and relationships stay with the learner. The reason is practical: those three operations are the part of the work whose value is rising.

How to practice this

1. Write down the artifact that should exist when the learning is done. Notice whether it is a thing that can sit on a screen or in a room. Notice whether making it leaves you a harder capability or only a faster version of the same work.
2. Ask the model for resources, prerequisites, and examples close to your artifact. Rank their importance yourself. Notice any urge to ask the model which ones matter, and do not act on it.
3. Before you study, get a study guide, a pre-quiz, and a list of key terms, and take the quiz. Notice that a poor score is fine. Stop priming when the topic has a big-picture structure and the main study feels like the next step.
4. Decide what matters, what groups together, and what is unclear before asking the model to organise anything. Then draw how the pieces connect yourself, badly if needed, and let the model critique it. Notice if the drawing will not start; the structure is not in your head yet.
5. Start the artifact while you are still working through the material. Every five to ten minutes, think of one place the new information applies. Notice if you cannot explain, modify, defend, or apply the artifact; the loop is not finished.
6. While the session is open, write the hardest applied question the material can bear. Notice if no hard question comes while found questions are easy; that is a conceptual gap. At the end, ask whether the model sped up the learning or did the learning.

Related pages

• The Right vs Wrong Way to Work With AI: the same rule stated as a prohibition. Never ask for the answer; ask for the information that helps figure it out.
• Don’t Outsource the Learning: the evidence page. The same tool, used in two different ways, produced different comprehension.
• Prestudy: the priming step worked out, with the completion test for priming.
• Deep Processing: the dimension that the rule against handing over schema formation protects.
• ICS System: learning engineers design a personal system rather than hunting one technique.

6 more related pages

• The Shortcut Problem: the general mechanism. This workflow is a special case of it.
• Problem-First Learning: problem before tool; the engine on the bench.
• Layers of Learning: logic, concept, and detail, presented mixed together in sources.
• Learning Styles Myth and Multimodal Learning: matching a declared style produces no learning advantage.
• Spaced Interleaved Retrieval: the real interleaving effect, inside a domain.
• Interleaving for Complex Problem Solving: reconstruction under changed variables.

Open questions

Is a first-attempt rule enough to distinguish clarification from outsourcing, or does the type of question matter?

Does a second-medium pass preserve engagement when the reason is dead time or a second encounter, not a declared style?

How should this framework adapt for exploratory learning where no artifact is known in advance?

Which steps should be automated as a repeatable skill, and which must stay hand-run?

Which of the three in-session signals fails first?

Sources

• How I Learn Things Really Fast (with AI)
• How to Learn ANYTHING Faster Than Everyone
• How To Become A Top 1% Learner (Without Being Smart)
• Elon Musk - How To Learn Anything
• Pashler, H., McDaniel, M., Rohrer, D., & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119.
• Richland, L. E., Kornell, N., & Kao, L. S. (2009). The pretesting effect: Do unsuccessful retrieval attempts enhance learning? Journal of Experimental Psychology: Applied, 15(3), 243–257.
• Kornell, N., Hays, M. J., & Bjork, R. A. (2009). Unsuccessful retrieval attempts enhance subsequent learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(4), 989–998.