What Makes Content Adaptive?

There is a common misunderstanding that adaptive content is content that gets generated on the fly. It is not. Adaptive content describes two things working together:

• The properties of the content itself, the structural choices that let a system reason about it.
• The clarity of the learning objectives it is organised around, because the objective is the unit the engine tracks proficiency against, recommends towards, and reports on.

The technology is only as good as the content underneath it, and especially the learning objectives that organise that content. An adaptive engine cannot personalise its way out of vague, compound or unmeasurable objectives.

You cannot bolt one onto the other later.

Five qualities decide whether a piece of content behaves well in an adaptive system. Most teams already do some of this implicitly. The shift is doing it deliberately and consistently.

If there is one place where content teams get the most leverage, it is here. Everything else in this guide is downstream of how well your learning objectives are written.

Get the vocabulary straight

The terms in this space are used loosely, which is part of why teams talk past each other. Three layers, three jobs:

• Goal: the broad ambition (e.g. “Develop algebraic reasoning”)
• Outcome: what is measured at unit or course end (e.g. “Solve linear equations in one variable”)
• Objective: what the learner can do at the end of one lesson (e.g. “Isolate a variable in a two step linear equation by applying inverse operations”)

The objective is the unit an adaptive system actually tracks. Get this layer right and the rest follows.

Start with backward design

Before writing any learning objective, agree the assessment that would prove a learner has mastered it. This is the core of backward design, popularised by Wiggins and McTighe and now standard practice across the field. Backward design is a three stage approach: define desired results, determine acceptable evidence of learning, and only then plan learning activities.

In adaptive content, this matters twice over: the objective drives the assessment, and the assessment produces the data the engine learns from. If you cannot picture the question that would prove mastery, the objective is not yet ready to write.

Use the A-SMART formula

Recent work has refined the classic SMART rubric into “A-SMART”, emphasising that outcomes should begin with an action verb to make them immediately measurable: Action oriented, Specific, Measurable, Attainable, Relevant, Time bound. A useful supplementary structure is the ABCD model: Audience, Behaviour, Condition, Degree. Most well written objectives implicitly contain all four.

The second one tells you, the engine, and the learner exactly what success looks like.

Pick verbs that can be observed

Action verbs do most of the work in a learning objective. The single most common defect in legacy objectives is the use of verbs that cannot be observed or graded. Avoid verbs like “understand”, “appreciate”, or “know” and replace them with verbs you can measure:

If you cannot picture the test question that would measure the verb, change the verb.

Use Bloom’s revised taxonomy to set the level

The revised Bloom’s Taxonomy uses active verbs (Remember, Understand, Apply, Analyse, Evaluate, Create) to clearly define measurable learning objectives. For an adaptive product, Bloom’s gives you two things at once: a consistent level tag for every objective that becomes useful metadata, and a check on whether your instruction matches your assessment. If the objective is at “Apply” level but every assessment item is “Remember”, you are not measuring what you wrote.

One objective, one outcome

Compound objectives are a quiet killer. “Identify and explain the causes of the French Revolution” looks like one objective; it is two. Identifying is at Remember level; explaining is at Understand level. A multiple choice question can measure the first; a short answer response is needed for the second. Compound objectives produce muddled proficiency estimates because a learner can succeed on one half and fail on the other and the engine cannot tell which.

The right grain size: lesson scope, not unit scope

There is a temptation to write objectives at the unit or chapter level, partly because that is where syllabus documents live, partly because writing many lesson level objectives is more work. Resist it. Lesson scope is the level at which instruction and assessment can both fit on the same page, the verb genuinely describes one observable behaviour, and the volume targets make sense. Unit scope objectives are still helpful as a navigation aid for the wider course, but they are not the unit the engine works against.

The pedagogical underpinnings worth knowing

The volume targets earlier are not arbitrary; they sit on top of three well established findings from cognitive science.

Cognitive load theory. Working memory has a fixed capacity. A single, focused objective per lesson keeps intrinsic load manageable and gives the learner space to process rather than merely receive.

The worked example effect. For novices exposed to information for the first time, worked examples are better than problem solving, though this effect reverses as experience grows. Per objective volume should include at least one worked example alongside practice questions. An adaptive engine that has both can serve worked examples to learners who are struggling and pure practice to learners who are progressing.

Retrieval practice and spacing. Decades of evidence support the testing effect: pulling information out of memory is more powerful than re-reading it, and spacing those retrievals over time consolidates learning more durably. This is the deeper reason the assessment volume target is closer to fifteen than five: it gives the engine enough material to space practice properly without showing the same item twice in quick succession.

Stable identifiers and a version rule

Every learning objective should have a stable unique identifier that lives for the life of the objective, and a written rule for when an edit is significant enough to count as a new objective rather than a revision. The engine attaches everything (proficiency, recommendations, analytics) to that identifier. Silently changing what an objective means while keeping the same ID poisons every metric attached to it. The rule does not have to be sophisticated; it just has to exist and be followed.

For every learning objective, aim for:

• At least 2 to 3 instructional items, ideally with different angles or formats: a worked example alongside an explanation, a short video alongside text
• At least 5 assessment items, the floor for any reliable proficiency estimate
• Closer to 15 assessment items for robust diagnosis without learners seeing the same questions repeatedly, and to give the engine room to space retrieval practice properly

If you cannot hit these targets across your library, that is a planning conversation, not a blocker. Some objectives matter more than others; some need more depth.

You do not need to invent your taxonomy from scratch. In most subject areas there is already a well established curriculum framework you can map your objectives against. Tag against:

• Cognitive level: Bloom’s revised taxonomy
• National curriculum: Common Core, the Australian Curriculum (ACARA), Cambridge syllabi, the CEFR for languages
• Subject specific: NGSS for science, AP frameworks, IB subject guides
• Question type: multiple choice, short answer, free response, drag and drop

The point is interoperability. The 1EdTech CASE standard facilitates the exchange of information about learning and education competencies between systems. The Quality Matters K-12 rubric is widely used to evaluate whether activities are clearly aligned to objectives and standards. Tagging against established frameworks means your metadata travels: between your products, between systems and between you and any partner you integrate with.

Three threads from current research are worth folding into how content teams plan.

The five properties cover how to structure individual objectives well. They do not cover how objectives relate to each other and for an adaptive engine, the relationship graph is where most of the intelligence lives. Once objectives are clean and discrete, the next question is: what does the engine need to know about how they connect?

The label trap

The most common mistake content teams make when approaching this for the first time is trying to solve the relationship problem by creating new types of objectives, “enabling learning objectives”, “sub objectives”, “foundational objectives”. This feels like progress because it produces a taxonomy. It actually makes the problem worse. Creating a new objective type collapses a relationship into a label. Once you call something an “enabling LO”, you have described what it does in relation to something else but you have not recorded what that something else is. The engine cannot traverse a label; it can traverse a link.

The Year 3 / 4 / 5 problem

Traditional content development for a given year level typically produces three content tiers: baseline, advanced, and easier remediation. In a non-adaptive product this works because a teacher selects the tier. In an adaptive product, the same three tiers correspond to something structurally different:

• The “easy” remediation content is often not easier content for the same objective; it is content aligned to a prerequisite objective that may sit in a prior year.
• The “advanced” extension content is often not harder content for the same objective; it is content aligned to an extension objective that may sit in a subsequent year.

When these are filed as difficulty variants of one objective, the engine cannot distinguish between a learner who has not yet reached the prerequisite and a learner who has mastered the current objective and is ready for extension. The fix is not to create three difficulty tiers within an objective. It is to write each objective at lesson scope for the year it belongs to, then record its prerequisite and extension relationships as explicit links.

A lightweight relationship model

Three relationship types cover the vast majority of cases. These should be recorded as explicit links between objective records, not as labels on individual objectives.

What this means in practice

Adding relationship mapping to the authoring workflow requires one additional step: after writing and validating each objective against the checklist, the author identifies and records any prerequisite or extension relationships to other objectives in the library. This is a content decision that belongs with the author, not with a separate metadata team.

• Prerequisite objective IDs
• Extension objective IDs

Where a prerequisite exists in a prior year level, record the link anyway, even if it points outside the current project scope. A broken link is better than a missing link; at least it is visible and fixable. Co-requisite relationships are lower priority to record explicitly, as the engine can often infer these from shared content alignment. Prerequisite and extension links should be the focus in the first pass.

Before declaring a set of learning objectives “adaptive-ready”, work through the following. These are the minimum structural conditions for a content library that an adaptive engine can work with.

If you can tick most of those, you are a long way ahead of where most libraries start.

1. Wiggins, G., & McTighe, J. (2005). Understanding by Design (2nd ed.). ASCD.
2. Anderson, L. W., & Krathwohl, D. R. (Eds.). (2001). A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom’s Taxonomy of Educational Objectives. Longman.
3. Li, H., Fang, Y., Zhang, S., Lee, S. M., Wang, Y., Trexler, M., & Botelho, A. F. (2025). ARCHED: A Human Centered Framework for Transparent, Responsible, and Collaborative AI Assisted Instructional Design. Proceedings of Machine Learning Research, 273, 94–104.