Instructional Scaffolding: Examples and How to Fade Support

Go from storyboard to an interactive course in minutes

Start Free

Picture a corporate travel consultant in her third week. In practice she has built several itineraries with hints on screen, and every live booking she makes still goes past her team leader before it is ticketed. Is she ready to work without the hints? Can the team leader stop checking? Nobody has written down what would show it, so the help either stays for good or comes off on a date someone picked at the start.

Deciding what help to give, and when to take it away, is what instructional scaffolding is about. I'm an instructional designer at Coursebox, where we make instructional design software that turns a brief into a course, so I'll mention our product where I tested it. Here I explain where the term comes from, the six things a scaffold does and how to fade each support on evidence, using one hypothetical travel consultant course, two real studies from health care and a free planner.

What is instructional scaffolding?

Instructional scaffolding is temporary help, built into a task or given by a person, that lets a learner complete work they cannot yet do alone, and that is gradually reduced as they succeed until they can do it without the temporary help, while keeping permitted job aids and required controls.

The term comes from a 1976 paper on tutoring by David Wood, Jerome Bruner and Gail Ross, The role of tutoring in problem solving. They described a "scaffolding" process "that enables a child or novice to solve a problem, carry out a task or achieve a goal which would be beyond his unassisted efforts" (p. 90). The tutor does this by "controlling" the parts of the task that are "initially beyond the learner's capacity", so the learner can work on the parts they can manage. In their study a tutor helped 30 children aged 3, 4 and 5 build a pyramid from 21 wooden blocks. Her aim "with every child was to allow him to do as much as possible for himself" (p. 92). She gave a spoken instruction first and stepped in more directly only when a child "failed to follow a verbal instruction".

Scaffolding in education now covers far more than one tutor and one child. Teachers use it for reading and writing, trainers use it for software and procedures, and course designers build it into self-paced e-learning as worked examples, hints and checklists. The idea stays the same in each setting: help that fits what this learner can do now, and that comes off as they get better at it.

Three features separate scaffolding from support in general. In their 2010 review of a decade of research, Janneke van de Pol, Monique Volman and Jos Beishuizen named them as "contingency, fading, and transfer of responsibility" (p. 271):

• Contingency: the help is adjusted to the learner's current level, which means you first have to find out what that level is.
• Fading: "the gradual withdrawal of the scaffolding" (p. 275), at a rate that depends on the learner's competence.
• Transfer of responsibility: as the help fades, the learner takes over more of the task until it is theirs.

Fading is part of the definition, not an optional extra. Roy Pea put it bluntly in his 2004 paper on the meanings of scaffolding: "If the support does not fade, then one should consider the activity to be distributed intelligence, not scaffolded achievement" (p. 432). Reading researchers P. David Pearson and Margaret Gallagher described the same hand-over in 1983 as "the gradual release of responsibility", in the technical report behind their model of reading comprehension instruction. The classroom shorthand "I do, we do, you do" sums up that release, but it does not appear in their work.

Wood, Bruner and Ross listed six functions their tutor performed (p. 98). They are the clearest answer to the question "what does a scaffold actually do?"

That pairing is useful, as long as each idea does its own job. Before you design any help you need to know what each learner can already do alone, which is the diagnostic question behind Vygotsky's zone of proximal development. This article picks up after that diagnosis: designing the help itself and taking it away.

Where to read about scaffolding: which source to start with

Scaffolding comes from papers rather than a book. Three of them cover most of what a designer needs.

• Wood, Bruner and Ross (1976). Read this first. It is 12 pages long and holds the original definition, the six functions on p. 98 and the rule that the learner's success or failure decides the tutor's next move.
• Van de Pol, Volman and Beishuizen (2010). Read this second. It is open access, reviews a decade of research, sets out contingency, fading and transfer of responsibility, and is honest about how hard scaffolding is to measure.
• Pea (2004). Read it if you design software or self-paced courses. It asks when help should fade at all, and it separates scaffolds that fade from supports that are there to help people do the work.

Types of scaffolding: hard, soft and the supports that stay

No agreed set of scaffolding types exists, whether the count is three, five or seven, and none comes from the people who defined the term. Three contrasts do come from the research, and each one changes a design decision.

• Hard and soft scaffolds. Thomas Brush and John Saye described hard scaffolds as "static supports that can be anticipated and planned in advance based upon typical student difficulties with a task" and soft scaffolds as "dynamic, situation-specific aid" that "requires teachers to continuously diagnose" (p. 2). A worked example, a hint ladder and a checklist are hard scaffolds you can build into a course. A team leader answering a question at the desk is a soft scaffold. Brush and Saye found hard scaffolds can work as "intermediate structures that support teachers in the task of soft scaffolding" (p. 10), so building them frees the person for the live help only they can give.
• Structuring and problematising. Brian Reiser (2004) described two jobs a scaffold can do. Structuring tries to "reduce complexity and choice by providing additional structure to the task". Problematising does the opposite: it points the learner at the hard part, which "may actually add difficulty in the short term, but in a way that is productive for learning". A hint that says "check the hotel cap" problematises; a form with the flights already filled in structures.
• Scaffolds that fade and supports that stay. Pea noted that many software supports "function not as scaffolds-with-fading but as scaffolds-for-performance" (p. 438), and that some could help experienced professionals too. For some supports, fading is "unnecessary and unproductive" (p. 442). A pre-flight checklist is not a scaffold the pilot is meant to outgrow.

Classroom strategies map onto the same ideas: modelling and think-alouds are demonstration, chunking reduces the degrees of freedom, questioning marks critical features, and a sentence frame is a hard scaffold that structures the task. Naming which one you are using tells you whether it should come off and how.

The six scaffolding functions, worked through one course

To show the six functions in use, I'll follow one hypothetical course from here on. A corporate travel management company in an Australian city has about 40 travel consultants who manage business travel for about 150 client companies. Five new consultants start together. Two come from leisure travel agencies, so they know airline booking basics but not corporate travel policies. Two come from a hotel front desk and an airline contact centre, so each knows one supplier's system well but not the full booking tool. One is a tourism graduate. All five have done the company induction and the booking tool's login and navigation module. None has built a corporate itinerary alone, and nothing in their earlier jobs shows they can apply a client's travel policy.

The task the course builds towards is this: from a traveller's trip request, build a multi-city business itinerary (flights, hotel and car) in the booking tool, apply the client's travel policy, flag anything outside policy for approval, and send the itinerary and quote within the client's standard of four business hours. The course uses one client, a mining services company, whose policy for this example says:

• economy class for flights under 5 hours;
• the lowest logical fare within 2 hours either side of the requested time;
• book at least 7 days before travel, or record a reason;
• preferred hotels only, with a nightly cap of $280 in Perth and $260 in Brisbane;
• a compact car unless three or more people travel together;
• anything outside policy needs the traveller's manager's written approval before ticketing.

The company also has its own 10-point booking checklist, from "read the whole request and confirm dates, cities and times" to "send the itinerary and quote and write the booking note". The policy settings and dollar figures are illustrative, not advice. By the end of week three, each new consultant should build and quote a multi-city itinerary inside this policy without hints and with no rework found by the team leader's check. Certification, authority to approve travel outside policy and advice on visas or entry rules are not outcomes of this course.

This example suits scaffolding because the task is a real procedure in a software tool, with judgement at the edges. It has an obvious worked example, a natural part-done version, rules that new starters miss, one support that should stay (the checklist) and one that should come off (the team leader checking every booking).

The six functions are not steps. A tutor uses whichever one the moment calls for, and a course can offer them in whatever order the task needs.

Recruitment

Recruitment means getting the learner interested in the task and committed to what it requires. Wood, Bruner and Ross's tutor had to "enlist the problem solver's interest in and adherence to the requirements of the task" before any building could start. In a self-paced course there is no tutor to do that face to face, so the task itself has to do the recruiting.

• In the classroom: a teacher opens a unit with a real letter from the local council asking students for their views.
• In the travel course: the course opens on a real-looking request from an impatient traveller who flies in 6 days, so the 7-day rule is in play from the first screen.
• Check it is working: learners can say what a finished booking includes before they start building one.

Reduction in degrees of freedom

This function simplifies the task by "reducing the number of constituent acts required to reach solution", so that the tutor "fills in the rest and lets the learner perfect the component sub-routines that he can manage" (p. 98). In course design it is often called a completion task: part of the work is done for the learner. The 4C/ID model uses completion tasks to fade support across whole classes of tasks.

• In the classroom: a maths worksheet gives the first three lines of a worked solution and leaves the last two for the student.
• In the travel course: on a Brisbane to Sydney to Brisbane trip of 2 nights, the flights are already built in the tool, and the learner adds only the hotel and car, applies the policy and writes the booking note.
• Check it is working: learners complete the remaining parts inside policy without asking how the pre-built parts were chosen.

Direction maintenance

Direction maintenance keeps the learner heading for the goal. Wood, Bruner and Ross described it as "keeping them in pursuit of a particular objective" and "making it worthwhile for the learner to risk a next step". Novices drift into the parts of a task they enjoy, or stall at the first obstacle.

• In the classroom: a writing teacher keeps a checklist of the essay's sections on the board and ticks them off as the class drafts.
• In the travel course: a progress cue shows which of the 10 checklist points are done, with a reminder of the four-hour service standard.
• Check it is working: learners finish practice bookings in full rather than stopping after the flights.

Marking critical features

Here the tutor "marks or accentuates certain features of the task that are relevant", and interprets the gap between what the learner did and what was needed: "His task is to interpret discrepancies". This is the function a hint usually serves. A good hint points at the feature, not at the answer.

• In the classroom: a science teacher asks "what changed between these two results?" instead of naming the variable.
• In the travel course: the hints flag the 7-day rule and the $280 Perth hotel cap, the two things new consultants miss most often in this example.
• Check it is working: learners spot the out-of-policy item on the next trip before opening a hint.

Frustration control

Wood, Bruner and Ross put it simply: "Problem solving should be less dangerous or stressful with a tutor than without". A learner who is afraid of breaking something stops experimenting, and one who is stuck for too long gives up.

• In the classroom: a teacher lets students redraft an answer after feedback without losing marks.
• In the travel course: a practice mode in which a wrong fare can be undone at no cost, and a hint that is always one click away while hints are still on offer.
• Check it is working: learners try a decision before opening a hint, rather than opening every hint first.

Demonstration

Demonstration shows the task done well. Wood, Bruner and Ross said it involves "an 'idealization' of the act to be performed". In e-learning this is the worked example, the usual starting point for novices in cognitive load theory.

• In the classroom: a teacher writes a paragraph live on the board, thinking aloud about each choice.
• In the travel course: a senior consultant builds a Brisbane to Perth return of 3 nights, booked 10 days out, on screen, explaining why at each step: why not the cheaper 6 am fare with a 9-hour wait, and why the $265 hotel rather than the $310 one.
• Check it is working: learners can answer "Which checklist step stopped the $310 hotel?" after watching.

How the six functions relate: contingency, fading and transfer

The six functions are not six steps, and they do not run in order. Wood, Bruner and Ross's tutor used whichever function the moment called for, and the deciding factor was the child: "The child's success or failure at any point in time thus determined the tutor's next level of instruction" (p. 92). That is contingency. Support goes up when the learner struggles and comes down when they succeed, which is why van de Pol and colleagues treat contingency, fading and transfer of responsibility as one package rather than three separate tips.

The authors also said a tutor needs two theories: "One is a theory of the task or problem and how it may be completed. The other is a theory of the performance characteristics of his tutee" (p. 97). For a designer that means knowing the booking procedure well enough to pick the hard parts, and knowing the five new consultants well enough to predict where each will get stuck. Without the second, they warned, "To the extent that the learner is at sea, so too is the tutor".

Even a trained tutor finds contingency hard to keep up. Wood, Bruner and Ross's tutor followed her rules for helping in 86% of 478 opportunities, and with the 4-year-olds she tended "to offer more help than allowed by the rules (27 out of 36 cases)" (p. 96). Over-helping is easy, which is why the travel course writes down when each support comes off instead of leaving it to the team leader's instinct.

In the travel course the functions overlap at every stage. The worked example is demonstration, but it also recruits, because it shows what a finished booking looks like. The hints mark critical features and control frustration at the same time. As the consultants succeed, optional help can decrease and their responsibility can grow. Restore suitable help if a new difficulty appears.

How to fade a scaffold

Fading is the part of a scaffolding plan that is easiest to leave vague. "Remove support gradually" is not a plan, and I could find no source for timetables such as "remove scaffolds after four to six weeks". Van de Pol and colleagues tie the rate of fading to "the child's level of development and competence" (p. 275), and they note that "several observations must be conducted over time to establish that fading has indeed taken place" (p. 287). So the plan has to say what evidence takes each support away. Here is the process I recommend.

1. List every support in the course. Include the ones nobody designed, such as a colleague who always answers questions, and note which of the six functions each one serves.
2. Mark each support as hard or soft. Fixed hints and worked examples are planned supports; a coach can respond to the current attempt. Adaptive software can also adjust help when it has reliable learner evidence and tested rules. Check which behaviour your course actually provides.
3. Decide which supports stay. If the job itself uses a support, such as a checklist or a policy lookup, it is a job aid, and it should stay. Mark it as staying so nobody strips it out for the sake of fading.
4. Put the rest in a fade sequence. A common sequence is a full worked example, then a completion task with part of the work done, then hints on request, then the job aid only, then the task done alone. Take away the demonstration first, then the part-done work, then the hints, then any reference aids that are not job aids.
5. Write the evidence that removes each support. Use things you can observe, such as correct attempts in a row without a hint, fewer hint requests, work done inside the standard or no rework found by a checker.
6. Remove one support, then check the next attempt. If the error comes back, put the support back for a set number of attempts. Contingency works in both directions.
7. Name the person who decides. Someone has to read the evidence and make the call. In a self-paced course that is usually a team leader or coach, not the course.

Fading works better when learners have to think about why each step is right, not only copy it. In Atkinson, Renkl and Merrill's 2003 study of faded worked examples, the abstract reports that fading alone "is not reliable on far-transfer tasks", while adding self-explanation prompts "produced medium to large effects on near and far transfer". In a classroom study with employees of an Australian financial services company, Kissane and colleagues (2008) reported in their abstract that gradually faded worked examples beat both example and problem pairs and plain problem solving on delayed and transfer tests.

The fade plan for the new travel consultants

Here is the sequence for the five new consultants. Each stage uses a different trip, so learners cannot simply repeat the last answer.

The 10-point checklist is: confirm dates, cities and times; open the client policy; check the 7-day booking rule; search within the 2-hour window; choose the lowest logical fare and record any exception reason; check connections and baggage; choose a preferred hotel within the city cap; add the car; list policy exceptions and request approval; send the quote and itinerary and write the note. The Sydney cap is not supplied, so the learner must obtain it before judging that booking. The policy owner must also explain how to judge a "logical" fare; a long wait alone is not a complete written rule.

Opening a hint is a reason to inspect the attempt, not proof of unreadiness. Check what the learner did and why, then provide another suitable trip or reduce help as the evidence supports. The two practice attempts and initial ten live bookings are illustrative planning budgets, not research thresholds. Vary the cases and judge the required actions, including when to ask for approval.

When to remove a scaffold: evidence, not dates

The team leader initially checks every live booking before ticketing. Ten bookings are an illustrative starting budget, not a validated readiness threshold. Reduce extra coaching only after observed, varied bookings meet the checklist and the learner can explain exceptions. If the team moves to one-in-five sampling, record which bookings were actually reviewed; ten bookings passing through the queue are not ten error-free observations. Keep ongoing risk-based checks and restore closer review when errors recur. Client approval rules remain in force throughout.

Taking help away on evidence is harder to run than taking it away on a calendar, because someone has to collect the evidence. The clearest real example I found of what happens when support is tapered by date comes from public maternity care in India.

Case study: Peer coaching for birth attendants in Uttar Pradesh, India

This was a large trial of a combined programme in public health care, and the tapering of support was part of the design rather than the thing being tested. In 120 government health facilities across 24 districts of Uttar Pradesh, birth attendants were not consistently carrying out essential birth practices. The BetterBirth programme, run by Ariadne Labs with Population Services International and the state government from December 2014 to September 2016, paired the WHO Safe Childbirth Checklist with peer coaching in 60 facilities and compared them with 60 matched facilities.

• What they did: Coaches visited twice a week for the first four months, weekly in months five and six, fortnightly in month seven and monthly in month eight, 43 visits in all, then stopped, as set out in a 2020 analysis of the coaching.
• Evidence and result: In the trial report, observed adherence to 18 essential birth practices was 73% against 42% at 2 months and 62% against 44% at 12 months, four months after coaching ended. By then, giving oxytocin soon after delivery, one of the essential practices, had fallen by nearly a third, from 79.5% to 53.9%. There was no difference in deaths and complications for mothers and newborns (15.1% against 15.3%, across 157,145 women).
• Limitation: The coaching came off by the calendar, not on each facility's performance, so the trial cannot separate fading from simply stopping. Adherence was observed in 4,888 births in a convenience sample of 15 matched pairs, and the checklist and coaching came as one package, so neither scaffolding nor fading can take the credit for the gap.
• Lesson: Monitor performance after support is reduced and retain a way to restore it. This trial did not compare calendar-based withdrawal with performance-based fading, so it cannot establish which fading rule works better.

For the travel consultants, reduce extra review from observed performance across varied bookings, keep the required approvals, and restore support when an error returns. Sampling provides evidence about the bookings checked, not every booking in the period.

The scaffold that stays

Not every support should fade. The 10-point booking checklist in the travel course comes off the screen at stage 5 but stays on the desk, because experienced consultants use it too. That is Pea's scaffold for performance rather than a scaffold with fading, and stripping it out to prove the course faded would make bookings worse, not better. Job aids like this belong to the moment of need at work, not to a learning stage people are meant to leave behind.

The test is simple. If an expert doing the job would still use the support, keep it and label it as a job aid. If only a novice needs it, plan its removal.

Scaffolding in self-paced e-learning, with no tutor in the room

Self-paced e-learning has a scaffolding problem built in: the tutor who noticed and adjusted is gone. Puntambekar and Hübscher's 2005 review of software scaffolds argued, in its abstract, that "ongoing diagnosis, calibrated support, and fading are being neglected" in tools, and Reiser noted "the limited ability of most scaffolded tools to individualize their support". His answer was to treat "learners, tools, and teachers" as a system. The course carries the hard scaffolds, and a person carries the soft ones.

A practical stand-in for contingent help in a course is a hint ladder: hints graded from a nudge to the answer, shown one at a time and only when the learner asks. The learner sets how much help they get, which is a rough version of what Wood, Bruner and Ross's tutor did. Here is the ladder for one decision at stage 3 of the travel course, where the traveller asks for a $320 hotel in Perth.

A ladder only works if the answer stays out of hints 1 and 2. If the first hint says "$320 is over the cap", every learner gets the full answer at the first click and the ladder does nothing. Write the nudge first, then check that someone who almost has it could finish from it without being told.

Hint use is not a diagnosis. If someone opens the answer, inspect their next attempt and ask them to explain the decision before concluding which knowledge or skill is missing. Fixed hint buttons do not adapt to performance automatically; an adaptive system would need dependable learner evidence and tested decision rules.

Help that stays on screen too long has a cost of its own. Slava Kalyuga and colleagues called it the expertise reversal effect: techniques that help novices "can lose their effectiveness and even have negative consequences when used with more experienced learners" (2003, p. 23). They also found that "a fading out procedure was superior to an abrupt switch" (p. 27). Without a tutor to take help away, a course has to build the removal into its structure, stage by stage.

On-screen prompts may also get in the way of the skill they support, as a small study of clinicians practising difficult conversations suggests.

Case study: Faded prompts for clinicians practising with virtual patients in Florida and New York

This was a small, non-randomised study in clinical mental health training. From 2020 to 2022, clinicians and clinicians in training at sites linked to the University of Florida and Mount Sinai in New York practised conversations with virtual patients at risk of suicide. In these conversations, how the clinician responds to emotion counts as much as what they ask.

• What they did: One yearly cohort of 27 got on-screen pop-up prompts based on an Acknowledge, Pursue, Confirm approach, and the next year's cohort of 35 got none. The prompts faded across sessions: a full explanation of the approach, then a short one, then a prompt without the approach named, then no pop-ups.
• Evidence and result: Yao and colleagues' 2024 study found a larger early improvement in verbal empathy with prompts, but that between-condition pattern was not detected at the third session. Later rating changes favoured the unprompted condition for likelihood of returning and satisfaction. This does not show that the prompted group lost its empathy gains, nor that removing prompts earlier would improve performance.
• Limitation: The two cohorts came from different years rather than random allocation, the 62 participants included students, the groups were small and the order of virtual patients was not counterbalanced. All outcomes were measured in simulation, not with real patients.
• Lesson: Check whether prompts support the target skill without disrupting the wider task. The study compared groups recruited in different years and did not test an early-fading rule. Try the placement of prompts and inspect both performance and learner experience.

The support question I get most often is about locks: how to lock the next module until a quiz is passed, or why a learner is stuck behind one. In the Coursebox support threads I reviewed from 27 January to 23 October 2026, 61 were about how learners are supported inside a course. Locking and sequencing was the leading theme, in 27 of them, split roughly between creators who wanted more gates and creators whose learners were stuck behind one. None of the 61 asked about hints, worked examples or job aids.

Scaffolding vs ZPD, gradual release, cognitive apprenticeship and 4C/ID

These ideas overlap, and they are often used as if they were one. Each answers a different question. The table is my synthesis, not a ranking.

Cognitive apprenticeship is the closest relative. It treats scaffolding and fading as one of its teaching methods, alongside modelling and coaching, and the cognitive apprenticeship model covers those methods and its own fading plan in depth. Pea credits Collins, Brown and Newman with the word "fading" itself, which does not appear in the 1976 paper. For the travel course, I would use cognitive apprenticeship to plan how the senior consultant thinks aloud in the worked example, and scaffolding to decide which hints exist and when each one comes off.

Criticisms and limits of instructional scaffolding

The criticisms fall into two groups: the term is used loosely, and the evidence behind it, especially for fading, is thin.

• The term is stretched. Pea warned that scaffolding "has become so broad in its meanings" that its significance is unclear (p. 423), and van de Pol and colleagues found it often "equated with (almost) any support given by the teacher" (p. 279). A worksheet that never changes is support, but not scaffolding.
• It can create dependency. Wood, Bruner and Ross named the risk themselves: "The major risk is in creating too much dependency on the tutor" (p. 98). Their own tutor gave the 4-year-olds more help than her rules allowed.
• The origin study is narrow. It involved 30 children aged 3 to 5 and one tutor on a block-building task. The six functions describe that tutoring, and they are not a validated design checklist.
• The evidence base is small and hard to measure. Van de Pol and colleagues found 8 effectiveness studies among the 66 articles they reviewed, mostly one-to-one tutoring on simple tasks, and called the evidence "somewhat limited" (p. 283). They found "no instrument in which the three key characteristics of scaffolding are considered together" (p. 286).
• Software scaffolds rarely fade. Puntambekar and Hübscher's abstract says diagnosis, calibrated support and fading are being neglected in tools, and Belland and colleagues' 2017 meta-analysis points the same way.

Fading is part of the definition, yet in a 2017 meta-analysis of computer-based scaffolding, fewer than one outcome in five involved any fading.

In Brian Belland and colleagues' 2017 meta-analysis, computer-based scaffolding had an overall effect of g = 0.46 across 144 experimental studies. Fading made no statistically significant difference, and the authors concluded that "fading as defined in the scaffolding literature is not necessary to promote student learning and performance". Those studies were in STEM education, mostly with students, and the paper flags that its adult estimate rests on very few outcomes, so none of it is a workplace figure.

So why fade at all? Pea asked the same question: "Why fade?" (p. 442). The better reasons are practical:

• Help that stays on screen becomes a cost as learners gain expertise.
• Gradual fading beat an abrupt switch in Kalyuga's research.
• Help that never comes off breeds the dependency Wood, Bruner and Ross warned about.
• The job will not provide hints.

In the travel course, the live booking tool has no hint button, and the team leader cannot check every booking forever. Fade the supports the job will not provide, keep the ones it does, and check after each removal.

What course requests leave out

To see whether people ask for scaffolding when they brief a course, I looked at a sample of course requests typed into Coursebox in the 90 days to 23 October 2026: one per person, from English-language browsers, with no uploaded file. Of 208 entries, 107 were course requests, and 43 of those were for workplace courses. Two independent AI readers each classified every entry against written definitions, and I re-read every positive by hand. Counts combine positives from either reader after manual review. They are coded estimates, not upper bounds: both readers may have missed requests. These briefs do not establish what the finished courses contained.

The one workplace request that asked for the task to be shown first wanted a demonstration video and step-by-step instructions, with no practice afterwards. Across all 107 requests, 2 were coded as asking for help during practice and none asked for help that fades. Where workplace requests mentioned support, it was the permanent kind, such as checklists or downloadable job aids (3 coded requests). Three were coded as asking for a final task done without help. The requests described the content and a test, and left out the middle: the worked example, the hinted attempt and the point where the help stops.

The sample is small and covers requests, not finished courses. Creators may add scaffolds later by editing, and a supervisor sitting with a new starter is scaffolding no prompt will show. Still, a course builder given a brief for content and a quiz will produce content and a quiz. If you want the help to fade, you have to write the worked example, the hints and the removal rules into the brief yourself.

Using scaffolding with an AI course builder

In late October 2026 I briefed Coursebox to build the travel consultant course. Coursebox is our product, so weigh my account with that in mind. The brief described the five new consultants, the mining services client's policy and the 10-point checklist, and asked for eight sections in order: "Before you start", a worked example, a completion task, three graded hints per decision, checklist only, on your own, the rules for when help comes off at work, and a wrap-up last. I asked for about 45 minutes and open-text answers.

Most of the fading came through. The first outline had the eight sections in the briefed order with the briefed titles. The builder asked one question my brief had not answered, "Where do the checklist and client policies actually live?", and I told it not to name a location and to tell learners to ask their team leader. It estimated about 37 minutes rather than 45, saying it had "prioritised meaningful practice and did not invent extra travel facts to pad". On the generated pages, "Watch it built" walks through the ten checklist steps one per card, with the briefed reasons: the 6 am fare "involves a 9-hour wait. It is therefore not the lowest logical choice", and the $265 hotel sits under the $280 cap. "Finish the booking" gives the flights as already booked and declined to invent a Sydney hotel cap the brief had not supplied. On "Hints if you need them", each decision has three hints behind separate "Show hint" buttons, and the answer appears only in the third.

The generated course still needed a clearer policy definition of "lowest logical fare" and the missing Sydney hotel cap before those decisions could be judged consistently. Its fixed hint buttons did not demonstrate performance-based adaptation. The revised live-review safeguards described here have not been retested in that course.

One gap undercut the fading rule. None of the practice pages kept the learner's work. "Hints if you need them" told learners "Your typed decisions are optional and are not saved", and nothing recorded which hints anyone opened. Yet the course's own rule says hints come off after two practice bookings in a row without opening a hint. The help faded on the page, but neither the learner nor the team leader could see how much help each person had used, and that is the evidence the rule for removing hints depends on.

I fixed it in the page editor. On the hints page I added a short-answer assignment under the hint activity, asking learners to record both decisions for the Melbourne to Perth to Brisbane trip, say why and say how many hints they opened for each, with a model answer. After I reloaded the learner view, the question, an answer box and a send button were there, and the model answer was saved in the editor. I checked the outline, the generated pages and the one block I added. I did not enrol learners or submit an answer, and a hint count that learners report themselves is only as good as their honesty. The generated page on removal rules states the same limit: "This does not mean the course tracks or enforces the process." Whether the help comes off at work is still the team leader's call.

Putting instructional scaffolding to work

Instructional scaffolding is help designed to come off, and the removal is the easiest part to leave out of a plan. Use it when a task is beyond new starters alone but within reach with help, and design the removal of each support at the same time as the support. The evidence that fading itself improves learning is thin, but the job will not give hints and a team leader cannot check every booking forever, so take each support away on evidence, check after it comes off and keep the supports the job itself uses.

Pick one task you train and list every support learners get on it. For each one, write down whether it stays or comes off, the evidence that will take it off and the error that will put it back.