Cognitive Apprenticeship Model: Six Methods and Examples

Go from storyboard to an interactive course in minutes

Start Free

Picture a trainee operator at a water treatment plant, three months into the job, on a night shift. On the trend screen, the filtered-water turbidity from one filter has started to creep up. The trainee passed the quiz on alarm levels and knows the procedure exists. What they do not know is what the senior shift operator would look at first, what he would rule out, or why he would run a jar test before touching the coagulant dose. He has twenty years of that thinking, and none of it has ever been said out loud.

I'm an instructional designer at Coursebox, where we make instructional design software that turns a brief into a course outline, so I'll mention our product where I tested it. Cognitive apprenticeship was built for this gap between knowing the rules and seeing how an expert thinks. Here I set out the model in the authors' own words, work the six methods and four dimensions through one water plant course, and add three real programmes with measured results and two free templates.

What is cognitive apprenticeship?

Cognitive apprenticeship is Allan Collins, John Seely Brown and Susan Newman's model of instruction for making an expert's hidden thinking visible: learners watch the expert reason through real tasks, then take the task over with coached support that fades until they work alone.

View the full-size diagram

Collins, then at BBN Laboratories, and Brown and Newman, of the Xerox Palo Alto Research Center, set it out in a 1987 technical report for the Center for the Study of Reading at the University of Illinois. The version usually cited is their 1989 chapter in a book edited by Lauren Resnick, Knowing, Learning, and Instruction: Essays in Honor of Robert Glaser. They started from a problem they saw in schools: "standard pedagogical practices render key aspects of expertise invisible to students" (1987, p. 2). In 1991 Collins, Brown and Ann Holum restated the model for teachers in the American Educator article "Cognitive apprenticeship: Making thinking visible", which gives the shortest definition: "Cognitive apprenticeship is a model of instruction that works to make thinking visible."

The name comes from traditional apprenticeship. The authors relied on Jean Lave's then unpublished study of "apprenticeship as practiced in a West African tailoring shop" (1987, p. 2), where apprentices learned through what Lave called observation, coaching and practice, or "what we, from the teacher's point of view, call modelling, coaching, and fading". The word "cognitive" marks the change of target. The focus is "on cognitive and metacognitive, rather than on physical, skills and processes" (p. 3): the reasoning, planning and self-checking that experts do in their heads.

The 1991 article names three differences from traditional apprenticeship:

• Thinking is made visible. "The teacher's thinking must be made visible to the students and the student's thinking must be made visible to the teacher." The authors call this "the most important difference".
• Tasks are set in meaningful contexts. In traditional apprenticeship, "the tasks come up just as they arise in the world"; the challenge is "to situate the abstract tasks of the school curriculum in contexts that make sense to students". The 1987 report adds that letting job demands pick the tasks is "one of the great inefficiencies of traditional apprenticeship" (p. 4).
• Skills are taught to transfer. The challenge is "to present a range of tasks, varying from systematic to diverse, and to encourage students to reflect on and articulate the elements that are common across tasks."

The model has four dimensions, which the authors said "constitute any learning environment": content, method, sequence and sociology. The six teaching methods are the best-known part. Here they are at a glance, in the wording of the 1991 summary table.

A programming course at the University of Helsinki shows the first three methods (modelling, coaching and scaffolding) clearly.

Case study: Learning to program at the University of Helsinki

The Department of Computer Science built this course explicitly on cognitive apprenticeship and combined it with practices from Extreme Programming, so the result belongs to the package. Its spring cohorts were mostly students taking computer science as a minor subject, and over the previous eight years only 43.7% of spring starters in the first programming course had passed.

• What they did: In spring 2010, lectures were cut from 5 to 2 hours a week and turned into worked examples that "demonstrate step by step how a solution could be devised for a problem" (modelling). Students did 15 to almost 40 small exercises a week, with the expected output shown as a built-in hint (scaffolding), in 8 hours a week of open labs where instructors "were not allowed to give direct solutions" and sent flawed work back (coaching). Articulation and reflection were not designed as separate methods.
• Evidence and result: In Vihavainen, Paksula and Luukkainen's 2011 paper, 70.1% of the 67 starters passed, against the long-term spring average of 43.7% and a previous best of 65.4%. In the follow-on course, 86.4% of 44 starters passed, against a spring average of 45.3%.
• Limitation: One cohort was compared with earlier years, with no randomisation and no significance test. The developers taught the course, and compulsory exercises raised the workload at the same time. The measure is the share of starters who passed, not programming at work. A later quasi-experiment with 124 students at a university in Istanbul also favoured the method, but its groups were not randomly assigned.
• Lesson: Move time from presenting content to coaching practice. Show a worked solution, set many small tasks, and keep a coach in the lab who questions the approach and sends flawed work back instead of handing out answers.

Collins, Brown and Newman's papers: which to read

The model exists in several versions, and the dates trip people up. Both 1987 and 1989 are correct: one is the technical report and the other the book chapter.

• Collins, Brown and Holum, "Cognitive apprenticeship: Making thinking visible" (American Educator, Winter 1991). Read this one first. It is free on the American Federation of Teachers website, written for teachers, and includes two think-aloud examples and a one-page summary of the framework.
• Collins, Brown and Newman, Technical Report No. 403 (January 1987). Read it for the full framework and the original evidence. It is free on ERIC and is the version to quote with page numbers. The 1989 chapter (pp. 453 to 494) is the version to cite for the model, but it is not free to read.
• Collins and Kapur, "Cognitive apprenticeship", in The Cambridge Handbook of the Learning Sciences (3rd edition, 2022). Read it for the current view. Collins wrote the first edition's chapter alone and Manu Kapur joined him for the 2014 and 2022 editions. The 2022 chapter's summary adds productive failure, where learners first try solutions that are wrong and are then guided towards correct ones.

The title wording varies too. The 1987 report says "the Craft" and the publisher's record of the 1989 chapter says "the Crafts", and both appear in good sources. A reference for the chapter in Harvard style:

The worked example: trainee operators at a water treatment plant

Here is the hypothetical example I'll use from here on. A regional water utility runs one conventional treatment plant that supplies about 40,000 people. Raw water goes through coagulation, flocculation and settling, then four dual-media filters and disinfection. The learners are trainee treatment-plant operators in their first six months. They hold an entry-level water operations qualification, have done the plant and safety inductions and can read the SCADA screens (the plant's control screens that show live trends and alarms). None has handled a water-quality excursion on shift, and the qualification does not show that they can.

The expert is the senior shift operator, who has worked at the plant for about 20 years. The shift supervisor is on call at night, and only the water quality manager contacts the health regulator.

The task is a rise in filtered-water turbidity on a night shift. The operator has to work out the likely cause, such as a filter near the end of its run, a change in raw water after rain, a coagulant dose problem or a filter ripening after a backwash. Then they take the first actions the plant procedure allows, record them in the shift log and call the shift supervisor at the right time. The plant's settings are an individual filter alert at 0.2 NTU, an action level of 0.3 NTU sustained for 15 minutes, and a critical limit of 0.5 NTU, at which the filter comes offline and the shift supervisor is called straight away. Operators may change the coagulant dose only within the range on the current jar-test sheet. These settings are made up for this example and are not guidance for any real plant.

So that learners' decisions can be judged, the proposed course would need five records supplied or prepared by the utility: the SCADA trend export for the incident (raw water turbidity, filtered turbidity per filter, coagulant dose, plant flow and filter run time), the turbidity response procedure, the latest jar-test sheet, the shift log and the filter backwash record.

The old approach failed in a familiar way. Trainees read the procedure, passed a quiz on alarm levels and shadowed night shifts where nothing happened. When a real excursion came, they knew the numbers but not how the senior operator reads three trends at once, what he checks first, or why he runs a jar test before he touches the dose.

The intended outcome is that by month six a trainee can work through a turbidity excursion on the plant procedure with the shift supervisor on call, say why they chose each action and escalate on time. It is not certified competence, and it gives no authority to change limits or contact the regulator. The example suits cognitive apprenticeship because the hard part is invisible. Two operators can look at the same trend screen, and only the experienced one sees that the problem is the coagulant dose, not the filter.

The course uses four cases, each harder or more varied than the one before:

• Case A, modelled: filter 3 rises slowly towards 0.2 NTU after 70 hours in service while the other filters hold steady. The senior operator talks it through: the filter is near the end of its run, so he backwashes it early and logs it.
• Case B, coached with three hints: after heavy rain, raw water turbidity triples and all four filters creep towards 0.3 NTU. The coagulant is underdosed for the new raw water, so the operator runs a jar test, adjusts the dose within the approved range, calls the shift supervisor because the action level is near, and logs it.
• Case C, one hint: filter 2 spikes to 0.4 NTU straight after a backwash and falls over 20 minutes. The filter is ripening, so the operator sends it to waste or holds it until it settles below the alert level, leaves the dose alone and logs it.
• Case D, no hints: at 2 am the coagulant dosing pump fault alarm sounds and turbidity is rising on all four filters. After the course, at work, each trainee picks a past excursion from the plant's shift log and talks it through with the shift supervisor.

The six methods, worked through one plant course

Each method is a different teaching move. Here they are in Collins, Brown and Newman's order, with what each one looks like for the trainee operators. The order is not a fixed sequence of lessons.

1. Modelling

Modelling means letting learners watch an expert do the real task. The 1987 report describes it as "showing an expert carrying out a task so that students can observe and build a conceptual model of the processes that are required to accomplish the task" (p. 16). In thinking tasks, it adds, this "requires the externalization of usually internal (cognitive) processes and activities". A silent demonstration shows the actions. For the reasoning, the expert has to say it out loud.

• Sample goal: Trainees can describe what the senior operator noticed on case A, what he ruled out and why he chose an early backwash.
• How to assess it: Ask trainees to retell the four parts of his reasoning in their own words, and check each part against the recording.
• In the course: A recorded think-aloud of the senior operator working case A on the utility-supplied trend export, in four parts: what I notice, what I rule out and how, what I decide and why, and what I write in the shift log.

2. Coaching

Coaching means watching learners do the task and responding to what they actually do. In the report it "consists of observing students while they carry out a task and offering hints, scaffolding, feedback, modelling, reminders, and new tasks aimed at bringing their performance closer to expert performance" (p. 16). The 1991 article calls coaching "the thread running through the entire apprenticeship experience", so it does not stop after one module.

• Sample goal: On case B, trainees name the likely cause and their first action before opening any hint, then correct their plan after feedback.
• How to assess it: Read each trainee's written plan for case B against the procedure, note which hints they opened, and give feedback on the reasoning as well as the answer.
• In the course: Case B as an open-text task with feedback written by the senior operator. On shift, the shift supervisor coaches the same reasoning on real trends.

3. Scaffolding and fading

Scaffolding is the support that lets a learner do a task they could not yet do alone, and fading is taking that support away. The report defines scaffolding as "the supports the teacher provides to help the student carry out a task" and fading as "the gradual removal of supports until students are on their own" (p. 17). It also sets a condition: scaffolding needs "accurate diagnosis of the student's current skill level or difficulty". Making the tasks harder is a different idea: the sequencing principle of increasing complexity.

• Sample goal: Trainees work case C with one hint and case D with none.
• How to assess it: Track which hints each trainee opened on cases B and C. A trainee who needed all three hints on case B gets another coached case before case C.
• In the course: Three hints on case B (compare the filters with each other; look at the coagulant dose against the raw water trend; check what the jar-test sheet allows), one hint on case C ("when did the spike start?") and none on case D.

4. Articulation

Articulation means getting learners to put their own knowledge and reasoning into words. The report describes it as "any method of getting students to articulate their knowledge, reasoning, or problem-solving processes in a domain" (p. 17), and names three ways: inquiry teaching, thinking aloud while solving a problem, and taking the critic's role in group work. Once the reasoning is written or spoken, a coach can see it and correct it.

• Sample goal: Trainees explain in writing why they chose each action on cases B and C, using the trend data.
• How to assess it: Check whether each explanation names its evidence (which filters, which trend) and the procedure setting behind the action.
• In the course: An open-text box after each case: what do you think is happening, what would you do first, and why?

5. Reflection

Reflection means comparing your own reasoning with someone else's. The report defines it as "enabling students to compare their own problem-solving processes with that of an expert, other students, and ultimately, an internal cognitive model of expertise" (p. 17). The 1991 article suggests an "abstracted replay", a replay of a performance cut down to its key decisions, to make that comparison easier.

• Sample goal: Trainees identify one point where their reasoning on cases B and C differed from the senior operator's, and one check they will now make first.
• How to assess it: Look for a specific difference, such as checking the raw water trend before the filters, rather than a general statement.
• In the course: After case C, trainees see the senior operator's short comments on both cases beside their own answers, then write what they would now check first.

A trial with resident doctors in Chengdu used all six methods on real patients, with articulation through questioning in group discussion and reflection through weekly journals.

Case study: Intensive care resident doctors in Chengdu

The intensive care unit at the First Affiliated Hospital of Chengdu Medical College in China designed this teaching explicitly on cognitive apprenticeship, combined with case-based teaching and extra training for the supervising doctors. The authors' problem was that experienced doctors' reasoning is tacit knowledge that "learners cannot directly observe", and bedside teaching left residents short of clinical reasoning for complex cases.

• What they did: In 2024, residents on a two-month intensive care rotation were randomly split into two groups taught "the same curriculum with identical contact hours". In the experimental group, trained supervisors talked through real cases to make their reasoning explicit, gave feedback on daily rounds, handed over patient responsibility gradually, questioned residents in group discussion, ran weekly reflection journals and had residents set their own learning objectives.
• Evidence and result: In Zhang and colleagues' 2025 randomised trial in Advances in Medical Education and Practice, 115 residents were analysed (56 experimental, 59 control). On the Mini-CEX, an observed clinical skills rating from 1 to 9 scored by clinical instructors, the average rose from 4.29 to 7.07 in the experimental group and from 4.48 to 6.39 in the control group, a gain of 2.78 points against 1.88 (P < 0.001 for the difference). History taking, communication and clinical judgement improved more in the experimental group, but physical examination, a more procedural skill, showed no significant difference.
• Limitation: It was one hospital over two months. The instructors who scored the Mini-CEX were not reported as blinded, and only the experimental group's supervisors had extra training, which the authors suggest may lift teaching on its own. The outcome is rated performance during training, not patient outcomes.
• Lesson: The study found larger gains on several reasoning-heavy ratings, but no statistically significant difference on physical examination. That does not establish that the method has no effect on procedural skills. Use cognitive apprenticeship where the hard part is invisible thinking, and train the experts first, because the methods live in the supervisors more than in the curriculum.

6. Exploration

Exploration means "pushing students into a mode of problem solving on their own" (1987, p. 18). The report calls it "the natural culmination of the fading of supports", involving "not only fading in problem solving, but fading in problem setting as well". Learners decide what the problem is, as well as how to solve it.

• Sample goal: Trainees handle case D without hints and say what they would check, what they would do, who they would call and when.
• How to assess it: Judge the written plan for case D against the procedure, including whether the critical limit and the call to the shift supervisor are handled on time. The shift supervisor then reviews the shift-log talk-through.
• In the course: Case D with no hints, followed by a task at work: pick a past excursion from the plant's shift log and talk it through with the shift supervisor, using the same four parts as the think-aloud.

How the six methods fit together

The six methods are not six steps. Collins, Brown and Holum wrote that they "fall roughly into three groups", and each group has a different purpose:

• The core: modelling, coaching and scaffolding. These are "the core of cognitive apprenticeship, designed to help students acquire an integrated set of skills through processes of observation and guided practice."
• Articulation and reflection. These help learners "focus their observations of expert problem solving" and "gain conscious access to (and control of) their own problem-solving strategies."
• Exploration. It is "aimed at encouraging learner autonomy, not only in carrying out expert problem-solving processes but also in defining or formulating the problems to be solved."

Coaching runs from the first coached attempt to the last review. Articulation follows every case, because each one ends with an explanation in the learner's words, and reflection comes before the unaided case, while there is still time to change how learners reason. What does move in one direction is the support: it starts high and fades, and exploration is where it runs out. The authors were blunt about recipes: "We don't believe that there is a formula for implementing the activities of modeling, scaffolding and fading, and coaching."

For the trainee operators, each case repeats the same pattern with less help: attempt it, explain the reasoning, then move to a harder case, with a comparison against the senior operator's reasoning before case D.

In practice, scaffolding, reflection and exploration are the ones that go missing. In Stalmeijer and colleagues' 2009 focus groups with 21 sixth-year medical students, modelling, coaching and articulation were common, while scaffolding, reflection and exploration "were mainly experienced during longer clerkships and with one clinical teacher".

The four dimensions: content, method, sequence and sociology

The methods are one of four dimensions in the framework. The other three cover what to teach, in what order and in what kind of learning environment. Lyons and colleagues' 2017 review of health sciences articles that used the model found that applications "tended to focus on the methods dimension (e.g., coaching, mentoring, scaffolding), with some consideration for the content and sociology dimensions". The other dimensions are where most of the unused design advice sits.

Content: four kinds of knowledge

Domain knowledge is "the conceptual and factual knowledge and procedures explicitly identified with a particular subject matter" (1987, p. 14), which is what most courses already teach. The other three are what experts add:

• Heuristic strategies are the "tricks of the trade"; "they don't always work, but when they do, they are quite helpful" (p. 14). For the operators: compare the filters with each other before blaming one.
• Control strategies "control the process of carrying out a task" (p. 14). For the operators: knowing when to stop investigating and call the shift supervisor.
• Learning strategies are "strategies for learning any of the other kinds of content" (p. 15). For the operators: using past entries in the shift log to learn from excursions they never saw.

Sequencing: the whole task first, then harder and more varied

Three principles set the order of tasks (pp. 18 to 19):

• Global before local. In Lave's tailoring study, apprentices "learn to put together a garment from precut pieces before learning to draw and cut out the pieces themselves" (p. 19), which lets learners "build a conceptual map, so to speak, before attending to the details of the terrain". The plant course opens with the whole treatment train and the one question on every excursion: is the water I am sending out safe, and what do I do now?
• Increasing complexity. Tasks are ordered so that "more and more of the skills and concepts necessary for expert performance are required" (p. 18). The tailors' apprentices start with drawers, which have straight lines and few pieces, before blouses. Case A involves one filter; case B involves the whole plant and a dose change.
• Increasing diversity. Tasks require "a wider and wider variety of strategies or skills" (p. 19), so learners see which parts of the reasoning carry across. Cases B, C and D have different causes, so trainees cannot pass by repeating the case A answer.

The authors admitted a limit: "We doubt that it is possible to sequence skills and tasks such that they undergo a monotonic increase in complexity" (p. 18). Order cases roughly, and watch where learners struggle.

Sociology: the learning environment

The 1987 report lists five characteristics of the learning environment (p. 20): situated learning, a culture of expert practice, intrinsic motivation, exploiting cooperation and exploiting competition. On competition, the authors warned that comparisons "must be made not between the products of student problem solving, but between the processes, and this is rarely the case" (p. 22). The 1991 summary drops competition, renames the culture of expert practice "community of practice" and lists four items. When you cite the sociology list, say which version you are using.

Collins, Brown and Newman's "situated learning" item means learning through realistic tasks. Jean Lave and Etienne Wenger's 1991 book used the same phrase for a different idea, legitimate peripheral participation. For the trainee operators, situated learning means the cases use the plant's own trend exports and records. Cooperation can mean pairs of trainees comparing their case C reasoning before they see the senior operator's. A culture of expert practice means operators talk about how they read trends, which is what the shift-log task with the supervisor is for.

Reciprocal teaching and the other original examples

Collins, Brown and Newman built the model from three teaching approaches that already worked in reading, writing and mathematics.

Reciprocal teaching

Palincsar and Brown's reciprocal teaching centres on "modelling and coaching students in four strategic skills: formulating questions based on the text, summarizing the text, making predictions about what will come next, and clarifying difficulties with the text" (1987, pp. 5 to 6). "It is called Reciprocal Teaching because the teacher and students take turns playing the role of teacher." The teacher models the four strategies, then hands the teacher's role to students while coaching them, so modelling, coaching and fading happen in one routine.

The early results were large. In a pilot with individual students who were poor readers, reading comprehension test scores rose "from 15% to 85% accuracy after about 20 training sessions", and six months later the students "were still at 60% accuracy" (1987, p. 6). The figures for the later group studies differ between the 1987 report and the 1991 article, so if you quote them, quote one version and say which. A later review gives a more sober picture.

Rosenshine and Meister's 1994 review of sixteen quantitative studies found the effect much smaller on standardised tests than on the researchers' own tests. The reported median is about a third of a standard deviation, but it is a different measure from the pilot's percentage accuracy gain, so the numbers cannot be compared directly. The gap between the two kinds of test is the practical lesson: check a course built on this model with tasks and measures from outside it, such as new cases at work, and not only with its own quiz.

Writing prompts and maths heuristics

Scardamalia and Bereiter gave student writers prompts, based on how expert writers plan and revise, to use as they wrote. Combined, the prompts produced "a tenfold increase in the frequency of idea-level revisions, without any decrease in stylistic revisions" (1987, p. 9). Schoenfeld taught college students heuristic strategies for maths problems and "found that learning these strategies significantly increased students' problem-solving abilities" (p. 10), with no figure given. The 1991 article reprints one of Schoenfeld's think-alouds as a sidebar, and it is the best short model I know of what an expert recording should sound like.

Cognitive apprenticeship in e-learning: what each method looks like online

Collins saw technology as a way to make the model affordable. Apprenticeship methods survive, the 1987 report says, in "human-resource-intensive settings, such as tennis coaching, learning foreign languages at Berlitz, or receiving training in medical diagnosis". Computer systems "need not take on the total responsibility. Instead, they only need to augment the master teacher" (p. 23).

That still holds. A self-paced course can model and scaffold well, and it can collect articulation. Coaching is harder, because coaching means observing learners "while they carry out a task", and a course sees only what learners type. Dennen and Burner's 2008 review of the research makes the same point: computers cannot adjust to learners' needs "in as subtle and personalized of a manner as a teacher might" (p. 434). Here is how the methods split between the course and a person.

In practice, the online version is the expert's recorded reasoning plus a planned handover to a person. A face-to-face programme for science teachers in Colorado shows how much of the modelling video can carry, with people doing the coaching.

Case study: Video-based coaching for science teachers in Colorado

BSCS Science Learning designed this year-long professional development programme for teachers of 4th and 5th grade explicitly on cognitive apprenticeship and situated cognition, and combined it with video analysis, a content framework, lesson plans and science content. The developers' problem was that these teachers are "typically reluctant to teach science" and rarely get training that changes what happens in class.

• What they did: After a two-week summer institute, teachers met monthly in study groups of 5 to 10 for 3.5 hours. They analysed video of more experienced teachers, with leaders modelling how to analyse it, then taught provided model lessons and analysed video of their own and their peers' lessons with feedback. In the second half of the year, "scaffolding supports such as model lesson plans are removed (fading)" and teachers planned lessons in a new content area. Leaders who had not developed the programme ran it in schools along Colorado's Front Range.
• Evidence and result: A cluster randomised trial covered 77 schools, 137 teachers and 2,823 students, and the comparison group got a content-focused programme of the same length. Roth and colleagues' 2017 design paper reports effects on teachers from 0.66 to 2.05 (2.05 for teaching practice) and a student effect of 0.68 on a science test. A later paper from the same team cites the published trial at 0.52 for students, so read the student effect as about half to two-thirds of a standard deviation.
• Limitation: The developers' team ran the research, the student test was written by the researchers, and the programme was a package, so the result belongs to the whole of it. A conference report on the trial found "only small and non-significant effects of these teacher variables on student achievement", so the gains in teacher knowledge did not on their own explain the students' gains.
• Lesson: Sequence professional development as watching experts on video, teaching with a script and getting feedback on your own recorded practice, then dropping the script and planning your own. Against an equally long content-only course, the programme as a whole changed teaching practice.

How to use cognitive apprenticeship to design a course

Collins, Brown and Holum summed up the teacher's job in three parts. To turn traditional apprenticeship into cognitive apprenticeship, teachers need to:

• "identify the processes of the task and make them visible to students";
• "situate abstract tasks in authentic contexts, so that students understand the relevance of the work";
• "vary the diversity of situations and articulate the common aspects so that students can transfer what they learn".

The workflow below is my own, built on that summary and carried through the plant course.

1. Check that the task needs it. Choose a complex task where experienced staff make decisions novices cannot see. If the task is rote, a clear procedure and practice will do.
2. Name the expert and book their time. Book the senior operator for the recording and the shift supervisor for coaching on shift. Without that time, the later methods do not happen.
3. Capture the expert's reasoning. Record a think-aloud on one real case in four parts: what they notice, what they rule out, what they decide and what they record.
4. Order the cases. Start with the whole job, then order cases by complexity and variety, from case A to case D, and use records supplied or prepared by the utility in each.
5. Write the hints and plan the fading. Write the hints for each case and decide what support each new case removes.
6. Add articulation to every case and reflection before the last. Ask for an open-text explanation after each attempt, and show the expert's comments beside the learner's answers before the unaided case.
7. Hand exploration to a person. End with a new case without hints and a real task reviewed by a supervisor, and give the supervisor a short coaching log.
8. Check with real tasks. Judge trainees on new cases and real shifts, not only on the course quiz. An observation checklist helps the supervisor review a real excursion.

The fading plan for the trainee operators

Here is the plan for the four cases. Each case removes some support, and each ends with a check before the trainee moves on.

What course requests leave out

To see how people brief courses for tasks like this, I looked at course requests typed into Coursebox in the 90 days to 7 October 2026: one per person, from English-language browsers, with no uploaded file. Of 266 prompts, 104 were course requests. Two independent AI readers each classified all 104 against written definitions of what each method would look like in a brief, and every positive call was re-read in full by hand. Counts combine the two readers' classifications and remain estimates. Checking positive calls does not establish how many requests were missed.

The nearest misses were a request to "Show correct vs incorrect techniques" in pictures and one for training videos for service plumbers that did not say the videos would show the work being done. The two requests for feedback on practice came from a personal study tutor and a make-up course with a final project rubric, and neither was a workplace course. In Collins, Brown and Newman's terms, the briefs in this sample asked for content to explain and left out the three methods they called the core. The sample is small and covers requests, not finished courses, and organisations may run demonstrations and coaching that never appear in a prompt. Still, if you want a course to carry an expert's thinking, you have to write the expert's reasoning, the coached attempt and the fading into the brief yourself, because a course builder cannot supply your expert's reasoning.

Cognitive apprenticeship vs situated learning, scaffolding and other models

Cognitive apprenticeship shares words with several neighbouring ideas, which is where most of the confusion comes from. The table is my synthesis, not a ranking. I found no study comparing them head to head.

The closest neighbour is situated learning. Brown, Collins and Duguid's 1989 article on situated cognition proposed cognitive apprenticeship, credited to Collins, Brown and Newman, as the teaching method to go with it. Lave's research on tailors fed both ideas. They still do different jobs. Situated learning explains why newcomers need access to real work, and cognitive apprenticeship tells the expert what to do once the newcomer is there.

Scaffolding has its own history. The term entered educational research through Wood, Bruner and Ross's 1976 paper on tutoring. Stone's 1998 review of the metaphor notes that it was proposed "without any explicit reference to Vygotsky", and that Cazden first made the link to the zone of proximal development explicit in 1979. Shvarts and Bakker (2019) have since traced the word to Vygotsky's notebooks and to Luria and Bernstein, but the concept as educators use it was developed in later tutoring research. The 1987 report cites neither Wood nor Vygotsky and uses scaffolding as an established term.

The zone of proximal development is Vygotsky's related idea about the gap between what a learner can do alone and with help. Cognitive load theory explains why novices need worked examples and why guidance should fade, and the 4C/ID model and Merrill's principles of instruction plan whole courses around real tasks. Social cognitive theory covers learning by watching models more generally. For the plant course, I would use cognitive load research to size the hints, and cognitive apprenticeship to decide what the senior operator records and what the shift supervisor coaches.

One more comparison clears up a common worry. Exploration can look like discovery learning, which Kirschner, Sweller and Clark's 2006 paper argues gives novices too little guidance. Their paper does not mention cognitive apprenticeship, and the model places exploration last, after modelling, coaching and scaffolding have done their work.

Limitations of cognitive apprenticeship

The model's limits come from its own sources as much as from its critics.

• It depends on an expert who can explain their thinking. The 1987 report itself allows that not all the strategies in complex work "can be captured and made explicit" (p. 16). An expert who has never explained a decision may need help to do it.
• It takes expert time and training. In Stalmeijer and colleagues' focus groups, the main problem was "variability in usage of the methods, which was attributed to teachers' lack of time and formal training". The Chengdu trial trained its supervisors before the rotation began.
• Scaffolding, reflection and exploration need continuity. In the same study, scaffolding, reflection and exploration came mainly in longer placements with one teacher, and shorter ones gave students mostly modelling, coaching and articulation.
• Most users adopt only the methods. The 2017 health sciences review found applications focused on the methods dimension, with some attention to content and sociology.
• Effects shrink on independent tests, and transfer is uncertain. Reciprocal teaching's median effect was 0.32 on standardised tests against 0.88 on the researchers' own tests. Dennen's 2004 review describes a study by Hendricks in which a cognitive apprenticeship group did better on the post-test but no better on a far-transfer task two weeks later.
• The evidence comes mostly from schools, hospitals and universities. I found no verified study of a private-sector employer's programme built explicitly on the model with a measured result, and each programme with good results combined the model with other elements.
• A self-paced course can coach only in part. Hints and written feedback are planned in advance, while a coach responds to what a learner actually does.

Using cognitive apprenticeship with an AI course builder

On 8 October 2026 I briefed Coursebox to build the plant course. Coursebox is our product, so weigh my account with that in mind. The brief asked for a self-paced course of about 45 minutes in eight sections: a short "Before you start", the whole job first, the senior operator's think-aloud on case A in the first person, case B with three hints, case C with one hint as its own "Less help" section, an "Explain and compare" section, case D with no hints plus the shift-log task, and a wrap-up as the very last section. It gave the plant's NTU settings and the rule that only the water quality manager contacts the regulator, and asked for open-text answers instead of quizzes.

Much of the first outline matched the brief. It opened with the whole job, with pages on the treatment train and the safety decision, and put the senior operator's reasoning on filter 3 before any practice.

The structure was where it slipped. The first outline came back with 5 sections and 12 pages instead of 8 sections. The "Less help" case had been folded into the comparison section, the wrap-up was missing and three practice items had come back as quizzes, so the drop from three hints to one to none no longer showed in the course structure. That drop is the fading, and fading is one of the parts of the model a self-paced course can carry well.

One follow-up message fixed it. I named the eight sections in order, said why "Less help" had to stand alone, asked for the wrap-up last and asked for open text in place of the quizzes. Then I reloaded the course, rather than trusting the reply, and checked it. It now had 8 sections and 20 pages in the brief's order, with "Less help: filter 2 after a backwash" as its own section and the wrap-up last, and the practice items were open-text responses with no quizzes left. In the generated pages, the think-aloud kept the senior operator in the first person with his reasons, starting "I notice Filter 3 first because it is the one that is changing", and then "I do not jump to a plant-wide cause just because one filter starts drifting up."

A few limits apply. The hints on case B appear as a list with an instruction to uncover them one at a time, rather than opening on a click. All hints are visible, so this does not control their release or show which hint was needed. Check every procedure figure yourself: one generated page mislabelled the 0.3 NTU action level, and I corrected it in the page editor. I checked the outline and the generated pages, and did not run the course with trainee operators. And the senior operator's reasoning came from my brief. A course builder can keep an expert's thinking and build the practice around it, but in real use that thinking has to come from your own expert.

With "Less help" back as its own section, the planned case sequence now steps down from a modelled case to three hints, one hint and none, The hint display still needs repair before it implements the staged access I briefed.

Putting cognitive apprenticeship to work

Cognitive apprenticeship answers a specific question: how do you teach a skill when the hard part happens inside an expert's head? Collins, Brown and Newman's answer is to make that thinking visible, let learners try the task with coaching and support that fades, and have them explain and compare their reasoning until they can frame and solve new problems alone. Use it for tasks that depend on judgement, not rote ones, and treat the six methods as moves that repeat on every case rather than as six modules. A course can carry the modelling, the hints and the comparison, but coaching on real work still needs a person with time to do it.

Start with one expert and one real case. Sit with them while they work it, write down what they notice, rule out, decide and record, and build your first coached case from that.