calendar
Edited on October 12, 2026

Instructional Design Models: How to Choose and Combine Them

Instructional design models and frameworks sorted by the problem they solve, with diagrams, a comparison table and a worked example combining three.

Carolina Martin, Learning Experience Manager at Coursebox AI
Carolina Martin
Customer Success Manager
Key takeaways

An instructional design model is a structured guide for one kind of design decision, such as how a training project runs, what happens inside a lesson, what learners must achieve or how you will judge whether the training worked.

  • Models are easier to choose by job, such as process models for the project and strategy models for the lesson.
  • ADDIE is not a one-way waterfall, because evaluation and revision can send a team back to any earlier phase.
  • Of 24 workplace course requests typed into Coursebox over 90 days, none named a model to plan the course with.
  • Choose one model for each design decision that is still open, which for many workplace courses means two or three.
  • Plan the check on behaviour at work, Kirkpatrick's level 3, before you build the course, because it shapes what the course must practise.
Which design decisions still need a model?

Five questions, about a minute. Answer for one course you are planning now, thinking about what is actually written down, not what you intend to do. This is a planning aid, not a validated measure or a certification.

1
2
3
4
5
Have you checked that training is the answer, rather than a job aid or a process fix?
Instructional Designers

Go from storyboard to an interactive course in minutes

Picture a new line lead at a wholesale bakery, a hypothetical one I'll use throughout this guide. On her first shift in the role she sat through a 30-minute slide deck on allergen law and passed a 10-question quiz. Then, under time pressure, she ran her first sesame-to-sesame-free changeover, skipped the second person's sign-off and checked the label on the first pack only. The course had covered the law but never practised the decisions she had to make at the line.

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. The usual advice for a problem like hers is to "use an instructional design model", but ADDIE, Gagné, Bloom and Kirkpatrick do different jobs. In this guide I sort about 25 models and frameworks by the problem each one solves, give each a short section, most with a diagram, and show how to combine the smallest useful set, with two real cases, a product test and two free templates.

What is an instructional design model?

An instructional design model is a structured guide for one kind of design decision, such as how a training project runs, what happens inside a lesson, what learners must achieve or how you will judge whether the training worked.

A six-row decision map connects unresolved design questions with possible frameworks: training need with action mapping or CRI; results and evidence with UbD and taxonomies; project organisation with process models; instruction with strategy models; workplace evaluation with evaluation frameworks; motivation with ARCS or SDT. These are starting points, not exclusive rules.

View the full-size diagram

Each model answers a different question. ADDIE organises the design functions of a project. Gagné’s nine events guide what instruction needs to accomplish, with events combined or repeated as appropriate. Bloom’s taxonomy classifies intended outcomes. Kirkpatrick helps teams define results and behaviour before training, then collect evidence during learning and in the workplace. These frameworks support decisions; none guarantees a successful course.

Three terms get mixed up:

  • Model and framework are used interchangeably in the field. Some writers keep "framework" for a broader structure, such as Understanding by Design, and "model" for a process, such as ADDIE, but I have not found an agreed line between them, so I treat them as the same thing.
  • A learning theory explains how people learn. Cognitive load theory, for example, says working memory is limited. A model turns ideas like that into design decisions.
  • A taxonomy is a classification. Bloom's and Fink's taxonomies sort learning outcomes; they do not tell you how to run a project.

Here are the model families in this guide and the question each one answers.

FamilyThe question it answersModels in this guide
Performance first and backward designIs training the answer, and what result and evidence do we design back from?Action mapping, criterion-referenced instruction, Understanding by Design
Design processHow will the project run, from analysis to evaluation?ADDIE, SAM, Dick and Carey, Kemp, rapid prototyping, Agile, Knirk and Gustafson
Classroom and lesson planningHow will one teacher plan a lesson and its media?ASSURE, Gerlach and Ely
Instructional strategyWhat happens inside the lesson, and in what order?Gagné's nine events, Merrill's First Principles, 4C/ID, elaboration theory, cognitive apprenticeship, Mayer's principles
TaxonomyWhat must learners achieve, and at what depth?Bloom's revised taxonomy, Fink's taxonomy
EvaluationHow will we know it worked?Kirkpatrick, CIPP, Brinkerhoff's Success Case Method
MotivationWhy will learners engage and keep going?ARCS, self-determination theory
note:Classroom models are process models in miniature

ASSURE and Gerlach and Ely focus on lesson planning, learners, media and classroom constraints. They overlap with wider design processes, but their emphasis and organisation differ. Use them when those decisions are your immediate problem.

Start with the design problem

Choose a model by the decision you have to make next, not by its popularity. Before opening any model, write down which of these questions is still open for your course:

  • Is training the answer at all, or is this a process, tool or job aid problem?
  • What must learners be able to do on the job, and what evidence would show it?
  • How will the project run, and who reviews what, when?
  • What happens inside the lesson, and in what order?
  • How will you know, weeks later, whether the work changed?
  • Why would learners engage with it?

Each open question points to one family. A question that is already answered needs no model.

People rarely start from a model, and I wanted to check that against real briefs. 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, and I re-read by hand every request that named a model and every workplace request. They agreed that 24 were clearly for people at work.

0 of 24
workplace course requests named an instructional design model to plan the course with, in the 90 days to 7 October 2026
Source: Coursebox course requests, read independently by two AI readers and checked by hand

The one workplace request that mentioned ADDIE wanted a course about ADDIE, not a course planned with it, and none of the 24 planned any check of what people did on the job afterwards. Most requests described a problem instead: 8 of the 24 were compliance or safety courses and 5 were onboarding, with some counted as both. With borderline requests counted, the workplace total rises to 27, and the count naming a model is still zero. The sample is small, it covers requests rather than finished courses, and teams may well plan evaluation outside the tool. What it does show is that, in this sample, people arrived with a problem, not a model, so the model has to be chosen from the problem and written into the brief on purpose.

Performance-first models: is training the answer?

Performance-first models start with what people must do on the job and ask whether a course is needed at all. Use them before any process model, when the brief arrives as "we need a course on X". The general method is a performance gap analysis, and for a whole team or role, a training needs analysis.

Action mapping

Action mapping, introduced by Cathy Moore in 2008, connects a measurable business goal to the actions people need to perform. Investigate with stakeholders why those actions are difficult: a tool, incentive or process problem may need a workplace change. If practice is part of the solution, design realistic decisions and supply only the information needed for them. See our action mapping guide and Moore’s explanation of the approach.

Action mapping connects a measurable business goal to required job actions. For each action, stakeholders investigate barriers and choose appropriate practice, task information or workplace changes. Branches show dependencies on the goal rather than a compulsory four-step content process.

View the full-size diagram

Criterion-referenced instruction

Criterion-referenced instruction comes from Robert Mager, whose 1962 book helped establish performance objectives, and later work with Peter Pipe on criterion-referenced instruction. Training is built from job-based objectives with conditions and criteria, each skill is checked against that standard, and people skip what they can already do. Use it when the job has a clear standard, as many safety and compliance tasks do.

Criterion-referenced instruction diagram: job analysis leads to performance objectives and matching skill checks, and an entry assessment sends each learner either to targeted practice or straight on when the standard is already met.

View the full-size diagram

Design process models: how the project runs

Design process models organise what gets analysed, designed, built, tried and revised, and who signs off. Name one overall project structure so responsibilities and approval points are clear. You can use prototype or Agile cycles within that structure; avoid duplicating documents and approvals without adding useful decisions.

ADDIE

ADDIE names five phases: analysis, design, development, implementation and evaluation. It has no single inventor: Michael Molenda's 2003 history of the ADDIE label found no original source for the acronym. The model descends from the Interservice Procedures for Instructional Systems Development (IPISD), written by Robert Branson and colleagues at Florida State University for the US armed services in 1975, and IPISD's fifth phase was called "Control", not evaluation.

ADDIE model diagram: analysis, design, development and implementation arranged in a loop around evaluation, which connects to every phase so feedback can send a team back to earlier decisions.

View the full-size diagram

warning:ADDIE is not a waterfall

ADDIE is often drawn as five boxes in a row, with evaluation at the end. The 1975 model already had feedback loops, and Molenda and Robert Branch both describe ADDIE as iterative, with evaluation and revision able to send a team back to any earlier phase. If your ADDIE plan only evaluates after launch, the problem is the plan, not the model.

SAM

The Successive Approximation Model (SAM) was published by Michael Allen with Richard Sites in 2012 as an iterative alternative to ADDIE. A project opens with a short collaborative "Savvy Start", then runs repeated small rounds of design, prototype and review. SAM1 is a single loop for small projects; SAM2 adds a separate development loop for larger ones. SAM makes early prototypes and repeated reviews explicit. ADDIE provides a broad structure that can also include prototypes, iteration and formal approval points. Choose according to the decisions and review capacity your team needs.

Successive Approximation Model diagram: a preparation phase leads into an iterative design loop of design, prototype and review, then an iterative development loop of develop, implement and evaluate, with SAM1 described as the simpler single loop.

View the full-size diagram

Dick and Carey

The Dick and Carey model, first published by Walter Dick and Lou Carey in 1978, is a systems model of about ten connected steps, from instructional goals through analysis, objectives, assessment and strategy to formative and summative evaluation. Its strength is alignment: the assessment instruments are written before the materials, and formative evaluation drives revision. It suits larger, high-stakes programmes where that rigour pays for itself.

Dick and Carey model diagram: instructional goals lead to parallel instructional and learner analyses, then objectives, assessments, strategy, materials and formative evaluation, with a revision loop back to earlier steps and summative evaluation kept separate.

View the full-size diagram

Kemp

The Kemp model, first set out by Jerrold Kemp and now maintained in Designing Effective Instruction (8th edition, 2019) by Gary Morrison and colleagues, arranges nine interdependent design elements in an oval rather than a line. A designer can start with whichever element is most pressing, such as learner characteristics or content, and revise the others as they go. It suits projects that arrive half-formed.

Kemp model diagram: nine interdependent design elements arranged in an oval around connected design decisions, surrounded by planning, revision and evaluation, showing there is no fixed starting point.

View the full-size diagram

Rapid prototyping

Rapid prototyping, described by Steven Tripp and Barbara Bichelmeyer in their 1990 paper on rapid prototyping for instructional design, has a team build a working prototype early, test it with learners and let what they learn reshape the analysis and design. The phases overlap instead of following each other. Use it when the riskiest part of a course is an interaction you cannot judge on paper, such as a branching scenario.

Rapid prototyping diagram: needs analysis, objectives, prototype construction, prototype testing and installation drawn as overlapping bands over time, with feedback from prototype use returning to analysis.

View the full-size diagram

Agile instructional design

Agile instructional design borrows short sprints from software teams: a prioritised backlog, usable pieces shipped often and a plan that changes with feedback. Megan Torrance's LLAMA (Lot Like Agile Management Approach) is a version built for learning teams. Agile is a way to manage the work rather than a theory of how to design it, so it pairs well with backward design, which sets what the sprints must produce.

Agile instructional design diagram: a prioritised learning backlog feeds repeated sprints of planning, building and checking, review and retrospective, with review feedback changing the backlog.

View the full-size diagram

Understanding by Design

Understanding by Design (UbD) plans the result and the evidence rather than running the project, so it sits alongside whichever process model the team uses. Published by Grant Wiggins and Jay McTighe in 1998, it plans backwards in three stages: the desired results, the evidence that would show learners have reached them, and only then the learning plan. The three stages describe the order of design decisions, not three sections of the finished course. Use it first when a brief lists topics but not what people must be able to do. Writing measurable learning outcomes is the first stage.

Understanding by Design diagram: plan backward from desired results, to the evidence that would show learners can do it, to the learning plan, keeping all three stages aligned.

View the full-size diagram

Knirk and Gustafson

Knirk and Gustafson’s 1986 model groups design work into problem determination, design and development. Establish the performance problem and learner starting point, specify objectives and instructional decisions, then develop, try and revise materials. The compact diagram is a high-level adaptation of the model’s three groups, not a complete task list or a claim that evaluation occurs only at the end. Our Knirk and Gustafson guide provides more context.

Knirk and Gustafson’s model groups work into problem determination, design and development. Problem determination includes the problem, entry skills, goals and organisation; design connects objectives, strategies and media; development includes materials, implementation, result analysis and revision. The high-level groups do not put evaluation outside the model.

View the full-size diagram

Classroom and lesson-planning models

These two models plan instruction for one teacher and one group, with the media and the room in view. They suit trainers who design and deliver their own sessions.

ASSURE

The ASSURE model was introduced by Robert Heinich, Michael Molenda and James Russell in 1982 for planning lessons that use media and technology. Its six steps run from analysing learners and stating objectives, through selecting and using methods, media and materials and requiring learner participation, to evaluating and revising. It is a good fit for a trainer planning a single session around a video, a simulation or an app.

ASSURE model diagram: six lesson-planning steps (analyse learners, state objectives, select methods and media, utilise media, require participation, evaluate and revise) with evidence feeding back into earlier decisions.

View the full-size diagram

Gerlach and Ely

The Gerlach and Ely model, from Vernon Gerlach and Donald Ely's 1971 book, sets content and objectives together, checks what learners bring, then plans strategy, grouping, time, space and resources as one block of decisions that affect each other. Evaluation feedback can revise any of them. The authors called it "a systematic, not a systems, approach", and it still reads well for classroom trainers who juggle rooms and timetables.

Gerlach and Ely model diagram: content and objectives specified together, entry behaviour checked, then strategy, grouping, time, space and resources planned together, with evaluation feedback able to revise any earlier decision.

View the full-size diagram

Instructional strategy models: what happens in the lesson

Strategy models shape the learning itself: what learners see, do and practise, and in what order. They pair with any process model.

Gagné's nine events of instruction

Gagné's nine events of instruction run from gaining attention, stating objectives and recalling prior learning, through presenting content, guiding learning, eliciting performance and giving feedback, to assessing performance and enhancing retention and transfer. Robert Gagné's The Conditions of Learning first appeared in 1965, but the nine events took shape in its 1970s editions; the current wording is in Principles of Instructional Design (5th edition, 2005). The events rest on a cognitive information-processing view of learning, not behaviourism, and they can be combined or repeated. Use them for a shorter, mostly linear lesson.

Gagné's nine events of instruction diagram: the nine numbered events grouped into preparing learners, supporting learning, and checking and transfer, with feedback able to lead to more practice.

View the full-size diagram

Merrill's First Principles of Instruction

Merrill's First Principles, published by M. David Merrill in his 2002 paper First Principles of Instruction, say learning is promoted when learners work on real-world problems and when new knowledge is activated, demonstrated, applied and integrated. That is five principles: the task at the centre and four phases around it. The phases can recur as tasks get harder. Use Merrill when the course builds a whole job task that involves judgement.

Merrill's First Principles diagram: activation, demonstration, application and integration arranged around real-world problems at the centre, with difficulty rising and support reducing across a progression of tasks.

View the full-size diagram

4C/ID

The four-component instructional design (4C/ID) model, developed by Jeroen van Merriënboer and set out in a 1997 book, trains complex skills through four components: whole learning tasks in classes of rising complexity, supportive information, just-in-time procedural information and part-task practice. The official 4C/ID site keeps the current account. It is the heavyweight option for complex skills that integrate several kinds of performance to build, such as diagnosis or troubleshooting.

Four-component instructional design diagram: whole learning tasks grouped into task classes of rising complexity with guidance fading in each, supported by supportive information, just-in-time procedural information and part-task practice.

View the full-size diagram

Elaboration theory

Elaboration theory, from Charles Reigeluth and Faith Stein in 1983, sequences content from a simplified representative whole task, concept or theory (the epitome) to more detailed layers, reconnecting each layer to the whole. Use it for a sequencing question: what to teach first in a large body of content.

Elaboration theory diagram: an epitome of the whole, elaborated through conceptual, theoretical or task sequences, with each new layer reconnected to the whole and reviewed with summarizers and synthesizers.

View the full-size diagram

Cognitive apprenticeship

The cognitive apprenticeship model, from Allan Collins, John Seely Brown and Susan Newman in 1989, makes expert thinking visible through modelling, coaching, scaffolding, articulation, reflection and exploration. These are teaching moves, not six fixed stages. It suits work where the hard part happens in someone's head, such as diagnosing a fault or reviewing a contract.

Cognitive apprenticeship diagram: four dimensions of a learning environment (content, method, sequence and sociology), with modelling, coaching, scaffolding, articulation, reflection and exploration grouped by purpose.

View the full-size diagram

Mayer's principles of multimedia learning

Richard Mayer's cognitive theory of multimedia learning holds that people learn more deeply from words and pictures than from words alone, within a working memory that handles words and pictures in separate, limited channels. His design principles, such as cutting extraneous material and placing words next to the graphics they describe, apply to every slide, screen and video in a course, whatever process or strategy model you use.

Mayer's cognitive theory of multimedia learning diagram: spoken words, printed words and pictures pass through ears and eyes into working memory, where learners select and organise verbal and pictorial models and integrate them with prior knowledge.

View the full-size diagram

Taxonomies: what learners must achieve

Taxonomies classify the outcomes you are aiming for, so you can check that objectives and assessments ask for the right kind and depth of learning. They do not run a project or sequence a lesson. Bloom's and Fink's are the two covered here. John Biggs and Kevin Collis's SOLO taxonomy (1982) is a third option, which grades the structure of a learner's answer from prestructural to extended abstract.

Bloom's revised taxonomy

Bloom's taxonomy began as a 1956 handbook by a committee of examiners, with Benjamin Bloom as editor. The 2001 revision by Lorin Anderson, David Krathwohl and colleagues crosses four types of knowledge with six cognitive processes, from remember to create, as David Krathwohl explains in his 2002 overview of the revision. Neither version was drawn as a pyramid, and the revision loosened the strict hierarchy. Classify the thinking an objective requires, not the verb it happens to use.

Bloom's revised taxonomy diagram: a grid of four knowledge types (factual, conceptual, procedural, metacognitive) against six cognitive processes from remember to create, used to place each objective, not as a lesson sequence.

View the full-size diagram

Fink's taxonomy of significant learning

Fink's taxonomy of significant learning, published by L. Dee Fink in 2003 and revised in 2013, names six kinds of learning: foundational knowledge, application, integration, the human dimension, caring and learning how to learn. Fink describes them as relational and interactive, not hierarchical. Use it when a course needs more than knowledge and skill, such as changing what people care about at work.

Fink's taxonomy of significant learning diagram: foundational knowledge, application, integration, human dimension, caring and learning how to learn arranged around significant learning, connected without a fixed order.

View the full-size diagram

Evaluation models: how you will know it worked

Evaluation models decide what evidence you will collect and when. Choose one at the start of a project, because the measures shape the design. Both published cases of deliberate model combination in this guide, from health workforce training, used one of these models to plan the follow-up before training began. General methods for evaluating a training programme apply to all three.

Kirkpatrick's four levels

The Kirkpatrick model evaluates reaction, learning, behaviour and results. Donald Kirkpatrick set out the four levels in a series of articles in 1959 and 1960, and James and Wendy Kirkpatrick's New World Kirkpatrick Model (2016) plans from results down, as Kirkpatrick Partners' current model page describes. The levels are not a causal chain: in Alliger and Janak's 1989 review, good scores at one level did not reliably predict the next. Return on investment as a fifth level comes from Jack Phillips, not Kirkpatrick.

Kirkpatrick model diagram: reaction, learning, behaviour and results drawn inside the performance environment, planned backwards from results, with a note that success at one level does not guarantee the next.

View the full-size diagram

Level 3, behaviour on the job, is a level that needs deliberate follow-up, because it happens weeks after the course and needs someone to do it. In one published train-the-trainer programme, planning it before the training gave a clear and uncomfortable answer.

51 of 79
lead public health nurses in Japan said they had run a role-play training session at their own branch within three months of a train-the-trainer day; 22 said they had not and 6 did not answer (self-reported)
Source: Arakida and colleagues, Nihon Koshu Eisei Zasshi, 2020

More than a quarter said they had not run a session, and the designers only know that because the follow-up was part of the plan.

Five women sit and stand around a wooden table with a laptop and an open notebook, one standing and talking while the others listen
RF._.studio / Pexels

Case study: Train-the-trainer for public health nurses in Japan

The Japan Health Insurance Association (Kyokai Kenpo), the public health insurer for employees of small and medium firms, needed lead public health nurses in its branches who could plan and run role-play training on health guidance for their colleagues.

  • What they did: With the International University of Health and Welfare, the association planned a 4.5-hour train-the-trainer day in August 2016 with ADDIE, and used Kirkpatrick's levels to set the evaluation, including a follow-up at three months. Level 4 results could not be obtained, so the authors used Level 3 behaviour evidence as a proxy. That does not make behaviour a measurement of organisational results.
  • Evidence and result: Arakida and colleagues reported in 2020 that "64.6% of participants held a role-play session at each branch within three months of the initial training", which is 51 of the 79 nurses. Self-rated knowledge and confidence rose sharply straight after the day and had slipped a little by three months, and knowledge fell significantly for nurses who had no previous facilitation training.
  • Limitation: This was action research with one group and no comparison group. Every measure was self-reported on named questionnaires in employer-run training, including the level 3 result, and knowledge and confidence were single items.
  • Lesson: An evaluation model can shape the design itself. Because the three-month follow-up was planned from the start, the result points to what to fix next: support for the nurses who did not run a session, and for those new to facilitation. A satisfaction score taken on the day would have shown none of that.

CIPP

The CIPP model, developed by Daniel Stufflebeam from 1966, evaluates context, input, process and product. Its purpose is to improve a programme while it runs, not only to prove it worked at the end, and the four parts can be used together or separately. Stufflebeam's CIPP evaluation checklist is the practical version. Use it for a programme that will run for years and needs decisions along the way.

CIPP model diagram: context, input, process and product evaluation around programme decisions, each answering a different question and usable together or separately.

View the full-size diagram

Brinkerhoff's Success Case Method

Brinkerhoff's Success Case Method, published by Robert Brinkerhoff in 2003, finds the most and least successful people who went through the training and interviews them to learn what helped or blocked results. It is quick and persuasive with managers, but Brinkerhoff himself cautions against using it to compute return on investment.

Brinkerhoff's Success Case Method diagram: plan the study, screen participants, interview the most and least successful cases, then verify their evidence and report what helped or blocked results.

View the full-size diagram

Motivation models: why learners will engage

Motivation models deal with whether learners pay attention, see the point and keep going. Add one when the audience is reluctant, such as staff sent to mandatory training.

ARCS

The ARCS model, set out by John Keller in his 1987 article on the ARCS model of motivational design, designs for four conditions of motivation: attention, relevance, confidence and satisfaction. Each has three design strategies, and the process starts by analysing the audience's motivation. It is not a four-step sequence. For the underlying theory of why people persist, self-determination theory explains motivation through the needs for autonomy, competence and relatedness.

ARCS model of motivation diagram: attention, relevance, confidence and satisfaction, each with three design strategies, around a motivational design process that starts with analysing the audience.

View the full-size diagram

Instructional design models vs learning theories

Learning theories explain how people learn, from behaviourism, cognitivism and constructivism onwards; models turn that into design decisions. Three theories come up often in workplace training design:

The 70-20-10 learning model is a rule of thumb rather than a measured ratio, but it is a useful reminder that a course is only one part of developing someone at work.

Five Moments of Need

Gottfredson and Mosher’s Five Moments of Need broadens the plan beyond a course: learning something new, learning more, applying knowledge, solving a problem and adapting to change. These are situations that can recur in any order. Plan training and support together, especially help at the point of application. Our Five Moments of Need guide introduces the framework; APPLY’s overview explains its performance-support emphasis.

Five Moments of Need presents New and More alongside Apply, Solve and Change. The panels describe recurring learning and work situations, not stages. Learning, practice and point-of-use support are planned together; the needs can occur repeatedly in different orders.

View the full-size diagram

Instructional design models compared

The table sets the main models side by side. It is my synthesis for choosing, not a ranking of effectiveness.

ModelFamilyThe question it answersBest forWeak spotPairs well with
Action mappingPerformance firstWhich on-the-job actions must change?Cutting a content-heavy brief down to decisionsNeeds access to the business goal and the jobAny process model, scenario practice
Criterion-referenced instructionPerformance firstWhat is the job standard, and who already meets it?Tasks with a clear standardHard to apply where the standard is a judgementGagné, Kirkpatrick level 3
ADDIEDesign processWhat phases will the project go through?Projects that need documented sign-offsEasily run as a slow one-way sequenceAny strategy model, Kirkpatrick
SAMDesign processHow do we iterate towards a good design?Uncertain designs and stakeholders who need to see draftsNeeds reviewers available for each roundAction mapping, Merrill
Dick and CareyDesign processHow do goals, assessments and materials stay aligned?Large, high-stakes programmesHeavy for a short courseBloom, Kirkpatrick
KempDesign processWhich design element needs attention now?Projects that arrive half-formedGives less guidance on orderAny strategy model
Rapid prototyping and AgileDesign processWhat can we build and test early?Risky interactions and fast-changing contentPrototypes can harden into the final productUnderstanding by Design, user testing
Understanding by DesignBackward designWhat result and evidence do we design back from?Briefs that list topics, not outcomesSays little about lesson sequenceMerrill or Gagné, Kirkpatrick
ASSURE and Gerlach and ElyClassroom planningHow will one trainer plan this session?Trainer-led sessions with mediaNarrow for multi-course programmesGagné, ARCS
Gagné's nine eventsInstructional strategyWhat should a lesson do, in what order?Short, mostly linear lessonsCan turn into a rigid checklistADDIE, Bloom
Merrill's First PrinciplesInstructional strategyHow do we build a whole task with judgement in it?Workplace tasks and decisionsNeeds realistic tasks to work withUnderstanding by Design, Kirkpatrick
4C/IDInstructional strategyHow do we train a complex skill over time?Complex whole-task competenceDemanding to designCognitive load theory
Bloom's revised taxonomyTaxonomyWhat kind and depth of learning is this objective?Writing and checking objectivesVerbs alone misleadAny process model
Fink's taxonomyTaxonomyWhich kinds of significant learning do we want?Courses aiming beyond knowledge and skillBuilt for college coursesUnderstanding by Design
KirkpatrickEvaluationWhat will we measure, from results down?Planning job and business measuresLevels are often read as a causal chainAny process model
CIPPEvaluationWhat should we decide about this programme now?Long-running programmesBroad, needs an evaluation ownerDick and Carey
Success Case MethodEvaluationWhat helped the best and worst cases?Quick evidence and stories for managersNot a measure of average effectKirkpatrick
ARCSMotivationWhy will learners engage?Reluctant or mandatory audiencesDoes not cover content or practiceGagné, Merrill

The model chooser template walks through each decision for your own course, with the bakery example filled in.

Instructional design model chooser

A worksheet that takes one course through each design decision in turn: is training the answer, what must learners achieve, how the project will run, what happens in the lesson, how you will know it worked and why learners will engage. For each decision you note whether it is already settled, the model you chose and why. Filled in for new line leads at a hypothetical wholesale bakery, with a blank version. No scores. Editable Word file, no email required.

How to combine models: the smallest useful set

A combination works when each model has one job and no two models fight over the same decision. These are the rules I use:

  1. Name the open decisions first. Write down which of the six design questions are still open, from whether training is the answer to why learners will engage, and cross out any your organisation has already settled, such as a fixed project process.
  2. Name an overall project structure. ADDIE, SAM, Kemp and Dick and Carey overlap in organising design work. Combine methods only when each adds a clear decision or useful review cycle, with one agreed set of responsibilities.
  3. Write the result and the evidence before the lesson. Understanding by Design does this job, with Bloom's or Fink's taxonomy to sharpen the outcome; action mapping and criterion-referenced instruction do it from the job standard.
  4. Choose the lesson model by the task. Merrill or 4C/ID for whole tasks with judgement in them, Gagné for a shorter, mostly linear lesson.
  5. Plan the evaluation at the start. Decide the level 3 check, who does it and when, before anything is built.
  6. Add a taxonomy or a motivation model only where a decision needs it. Bloom helps when objectives are vague; ARCS helps when the audience is reluctant.
tip:Keep the models out of the course

Models are for the design team. Learners should see a realistic task, practice and feedback, not four diagrams and a list of model names. If a model does not change a single decision in the brief, drop it.

A worked combination: allergen changeovers for new line leads

Here is the hypothetical bakery in full. The bakery supplies supermarket own-brand bread and buns from two production lines over three shifts. Some products contain sesame, and the rest are made on the same lines. Each year about 8 operators are promoted to line lead. They know the line, but their new job title does not mean they know the allergen changeover procedure, the swab test or how a line is released.

The task is to run a changeover from a sesame product to a sesame-free product and prepare its records for the QA release decision. The bakery's own procedure, which is this site's procedure and not regulatory guidance, has seven steps:

  1. Stop the line and remove all previous product, ingredients and packaging.
  2. Dry clean, then wet clean the listed contact points to the site clean schedule.
  3. The line lead inspects visually and a second person signs off.
  4. Swab test the 3 named contact points (seeder hopper, conveyor belt join and slicer guard); all must pass before restart, and a failed swab means re-clean and re-swab, never restart and monitor.
  5. Load the new label roll and check the first 5 packs against the product specification (product name, allergen statement and date code).
  6. Record every step on the changeover sheet.
  7. The QA lead releases the line.

Going the other way, from sesame-free to a sesame product, needs the label check but not the allergen clean. The old approach was a 30-minute slide deck on allergen law and a 10-question quiz on the first shift as line lead. New line leads passed the quiz, then skipped the second-person sign-off under time pressure and checked only the first pack's label. The intended outcome by day 30 is that changeovers are run and recorded to the procedure, with the sign-off and the 5-pack label check every time. Certification and authority to release a line are not outcomes; release stays with the QA lead.

Three complementary frameworks share evidence in a hypothetical bakery course: UbD plans the desired changeover result and evidence, Merrill shapes demonstration and practice, and Kirkpatrick plans behaviour checks and separate site-result tracking. Each required check is reviewed individually, observed readiness determines support and QA retains release authority.

View the full-size diagram

This procedure is invented to make the teaching example judgeable, not a validated food-safety plan. Real sites must use their approved allergen controls and authorised trial-pack and release procedures. Practice counts do not replace an observed readiness decision.

The example needs three models because it has three open decisions: what result and evidence count, what the lesson looks like, and how to check what a line lead does at the line weeks later. The bakery's usual review process is already settled, so no process model is added.

1. Understanding by Design sets the result and the evidence

The desired result is that a new line lead runs and records a sesame-to-sesame-free changeover to the fictional site procedure, ready for the QA lead’s release decision. Only the QA lead authorises release. The evidence is an observed changeover against the seven required checks, plus decisions on three changeover sheets in the course.

  • Sheet A: a swab failed at the slicer guard and the line was restarted "to monitor". Wrong: re-clean and re-swab before restart.
  • Sheet B: all swabs passed, but the label was checked on the first pack only and the second-person sign-off is blank. Wrong on two steps.
  • Sheet C: sesame-free to sesame, with no allergen clean required by this fictional procedure. The five-pack label check, all other applicable checks, records and sign-offs are complete, and QA approval is recorded. Correct for these supplied facts. Without those additional facts, the direction and label check alone would not establish that the changeover was ready.

Only then comes the learning plan.

2. Merrill's First Principles shape the lesson

The lesson starts from the whole task, a real changeover. Activation asks what goes wrong at changeover on these lines. Demonstration has a senior line lead run one, explaining each decision, especially the failed-swab rule and the 5-pack check. Application has the learner judge Sheets A, B and C with feedback, then begin with two supported changeovers with the senior line lead, with more practice as observed performance requires. Integration has the learner explain the failed-swab rule to an operator on their shift. If the course were a short refresher on the label check alone, Gagné's nine events would be a fair alternative; I chose Merrill because a changeover is a whole task with judgement in it.

3. Kirkpatrick level 3 checks the job

The QA lead first checks readiness before buddy support is withdrawn, then observes the first three changeovers without a buddy and reviews the next 30 days of sheets. These counts are illustrative planning choices, not evidence that three attempts establish competence. Use the same seven-point checklist, recording each applicable critical check separately; one missed release condition cannot be offset by other checks. QA retains release authority. Track allergen-related product holds and label errors as Level 4 results, with volume and other operational changes considered before attributing a change to training.

DecisionModelIn the bakery course
What must line leads do, and what shows it?Understanding by DesignDesired result; QA lead's 7-point checklist; judgements on Sheets A, B and C
What happens in the lesson?Merrill's First PrinciplesWhole changeover first, worked example, sheet practice with feedback, two supported changeovers, explain it to a shift
How will we know it changed the work?Kirkpatrick level 3QA checks readiness, observes the first 3 changeovers without a buddy, and reviews 30 days of records
How will the project run?Already settledThe bakery's usual review and sign-off process, so no process model is added

A real hospital ran a three-model combination of its own, with the measures fixed before anyone was trained.

A man in a blazer stands in front of a projected slide and speaks to a small audience seated on chairs in a training room
Pavel Danilyuk / Pexels

Case study: Lecture training for clinical teachers in Shanghai

Shanghai East Hospital, a national residency training base attached to Tongji University School of Medicine, had a lecture problem. In March 2019, residents voted the small lecture their least satisfying teaching activity, and 103 of 156 clinical teachers named lecturing as the skill they most needed to improve.

  • What they did: The hospital's teaching department ran the project with ADDIE from analysis to evaluation, built the course on Gagné's principles of instructional design and fixed a Kirkpatrick measure for each of the four levels before delivery. The course on five key lecture behaviours was revised five times over 10 trial lectures, then taught to 99 volunteer teachers in six sessions in June 2020. Each teacher then gave one real small lecture a month for four months.
  • Evidence and result: In Sun and colleagues' 2022 paper in BMC Medical Education, teaching secretaries who observed real lectures scored the five behaviours an average of 14.12 points higher out of 100 in October 2020 than in May, and a weighted performance value rose from 71.49 to 84.10.
  • Limitation: It was one group of volunteers at one hospital, with no control group. The raters knew the project, the weighting of the performance value was set after the data were collected, and scores from residents who had not been briefed were left out because they ran much higher. It measured observed teaching, not what residents learned.
  • Lesson: Each model did one job at its own decision point: ADDIE ran the project, Gagné shaped the lesson and Kirkpatrick fixed the measures, including observation of real lectures, before training began. The gain cannot be credited to the models alone, but the level 3 measure was usable because it was planned first.

Both published cases I found are health workforce training and both are single-group studies, so neither shows that a combination causes better results. I found no published corporate or government case of a deliberate model combination with a measured result.

Using a model combination with an AI course builder

On 12 October 2026 I briefed Coursebox to build the bakery course as this combination. Coursebox is our product, so weigh my account with that in mind. The brief set the desired result and the evidence first, asked for Merrill's order of whole task, then a worked changeover, then practice, and asked for a 7-point checklist for the QA lead's level 3 check. It named eight sections in order, gave the procedure and Sheets A, B and C, and asked for open-text questions instead of multiple choice.

Much of it came back as briefed. After one revision (the builder asked five confirmation questions, then ignored the answers until I listed the eight sections directly), the second outline had all eight sections in my order, 11 pages, no quizzes and five open-text assignments, with Sheets A, B and C on separate pages. On the rendered pages, the worked changeover came before the practice, and the QA observation checklist had seven cards, one per procedure step, plus the 30-day sheet review, kept as content for the QA lead rather than turned into a quiz. The Sheet A model answer had the failed-swab rule right: re-clean and re-swab, never restart and monitor.

The issue was Sheet C, the one sheet that was correct. Its task asked learners to "identify any issues" in 300 to 500 words. The hidden model answer that drives the AI feedback agreed that no allergen clean is needed going from sesame-free to sesame, then invented faults anyway, saying "several steps should have been outlined prior to executing the changeover" and listing a "Training Session" among them. A line lead who correctly judged the sheet as fine would have been marked against an answer saying it was not. That breaks backward design at the evidence: the task no longer tested the decision the desired result depends on.

I fixed it in the editor. The new Sheet C task, headed "Judge the sheet", gives the sheet's facts, asks the learner to say in a few sentences whether the changeover was run correctly and, if anything was wrong, what should have happened, and repeats that only the QA lead releases the line. The new model answer opens with "Sheet C is correct", explains why no allergen clean is needed in this direction, names the 5-pack label check it does need, and warns against demanding a clean that only the other direction requires. After a reload I checked both: "Sheet C is correct" was there, "Training Session" was gone and the word target had gone.

I checked the outline, the rendered pages and the model answers, not a run with line leads or a QA lead, and the estimate of about 1 hour 8 minutes is an outline figure, not a timing. A checklist on a page also proves nothing about whether anyone is observed. That part depends on the QA lead, which is why the level 3 check belongs in the plan before the course is built.

The written example above now supplies all applicable records and approvals for Sheet C. That clarification is an editorial repair to the example, not a new product test. Before using the generated course, check its task, model answer and release wording against those complete facts.

Model combination plan

A planning sheet for combining instructional design models: the desired result and evidence from backward design, the lesson order from Merrill or Gagné, and the level 3 and level 4 measures from Kirkpatrick, with who observes, when and with what. Filled in for new line leads running allergen changeovers at a hypothetical wholesale bakery, including the QA lead's 7-point checklist, with a blank version. No scores. Editable Word file, no email required.

Choosing your next model

Choose a model for the decision that remains unresolved. Taxonomies classify intended outcomes; process models organise design work; instructional strategies shape practice; evaluation frameworks help plan evidence. Combine them when each has a defined job. Plan the workplace check early so it shapes the practice, and add one later if it was missed, recognising that a late check cannot restore a missing baseline.

Your next step: take the course you are planning now and write one sentence for each of the six decisions. For each one that is still open, choose one model and record why. Start with the result, the evidence and the job check.

Frequently Asked Questions

The main types are performance-first models such as action mapping, design process models such as ADDIE and SAM, instructional strategy models such as Gagné's nine events and Merrill's First Principles, taxonomies such as Bloom's, evaluation models such as Kirkpatrick's, and motivation models such as ARCS. Each type answers a different design question, so they are usually combined rather than compared.

In practice, yes. The two terms are used interchangeably, although some writers keep "framework" for a broad structure such as Understanding by Design and "model" for a step-by-step process such as ADDIE. A learning theory is different: it explains how people learn, and a model or framework turns that into design decisions.

Corporate training usually needs a small set of models chosen by the problem: action mapping when the brief is a list of topics, a process model such as ADDIE or SAM to run the project, Merrill's First Principles for job tasks that involve judgement, and Kirkpatrick to plan the check on the job before the course is built.

ADDIE is still widely used, mostly as a shared language for project phases. It is not linear by origin: the 1975 IPISD model it descends from had feedback loops, and current accounts describe evaluation and revision feeding every phase. Teams that want shorter, more visible cycles often use SAM or Agile instead.

Yes, and most courses need more than one. Use one model per decision, such as Understanding by Design for the outcome and evidence, Merrill or Gagné for the lesson and Kirkpatrick for evaluation, and agree one set of responsibilities and avoid duplicate approvals when combining process methods.

Carolina Martin, Learning Experience Manager at Coursebox AI

Carolina Martin

Customer Success Manager

Carolina is a highly experienced learning designer with tech start-up experience. She is based in Sydney and previously designed courses alongside Coursebox founder Travis in a previous job. She is passionate about delivering engaging learning experiences and brings the same perspective to customer experience too. She manages our customer support documentation and product testing processes.