A Practical Workshop Framework
As organisations across higher education and publishing confront the challenge of integrating artificial intelligence into their operations, the question is no longer whether to adopt AI, but how to do so effectively. At Cambridge, we have been running an internal training programme focused on what we term “AI-ifying the organisation” – helping teams understand, adopt, and implement AI tools in ways that enhance rather than replace human expertise.
This article documents our approach to training publishing professionals in AI capabilities, drawing from a comprehensive workshop delivered on AI and Educational Publishing. Rather than treating AI as a mysterious black box, we focus on practical techniques that publishing teams can implement immediately, whilst also addressing the critical change management and upskilling requirements that successful AI adoption demands.
The framework we developed serves dual purposes: demonstrating what is genuinely possible with current AI technology, and providing a replicable training structure that other organisations can adapt to their own contexts. Throughout this article, we share specific examples, prompting techniques, and workflow implementations that have proven effective in real-world publishing scenarios.
Understanding AI and Large Language Models: Demystifying the Technology
Before publishing professionals can effectively use AI tools, they need a conceptual understanding of what these systems actually do, and crucially, what they cannot do. We begin our training by addressing the fundamental question: what exactly are Large Language Models (LLMs)?
The Calculator Analogy: Traditional Software vs AI
Traditional software functions like a calculator. When you input 2 + 2, you receive 4 every single time. The logic is deterministic, predictable, and perfectly reproducible. Educational publishers have relied on this type of software for decades – content management systems, typesetting software, and digital asset management tools all operate with this predictability.
AI systems, particularly LLMs, function more like weather forecasters. Given the same atmospheric conditions, a weather model might predict slightly different outcomes each time it runs, because it is processing probabilities rather than executing fixed logic. LLMs predict the most statistically likely next word based on patterns learned from vast training datasets, which means they can produce different outputs from identical inputs.
This fundamental difference has profound implications for publishing workflows. Where traditional software provides consistency and reliability, AI provides flexibility and generative capability but requires human oversight.
What AI Can (and Cannot) Do for Educational Publishing
Current AI systems excel at several tasks critical to educational publishing:
Content Generation and Adaptation: LLMs can draft curriculum-aligned content, adapt existing materials for different reading levels, and generate assessment questions based on learning objectives. They process natural language instructions and produce coherent, contextually appropriate text.
Translation and Localisation: AI can translate educational content across languages whilst maintaining pedagogical intent, though this requires careful structuring (as we discuss in the multilingual workflows section).
Content Analysis and Organisation: These systems can analyse large document collections, extract key concepts, identify alignment with curriculum standards, and suggest organisational structures.
Formatting and Restructuring: AI can transform content between formats such as converting prose into tables, generating summaries, or adapting content for different delivery platforms.
However, AI systems also have significant limitations that educational publishers must understand:
Hallucinations: LLMs will confidently generate plausible-sounding but entirely fabricated information. In one workshop demonstration, we asked an AI to write about South African author Deon Meyer. It correctly identified him as a crime fiction writer but then fabricated an entire biography with made-up book titles, awards, and biographical details. For educational content, where factual accuracy is paramount, this represents a critical risk.
Knowledge Cutoffs: Most LLMs have a training data cutoff date, meaning they lack knowledge of recent events, policy changes, or curriculum updates. A model trained on data through January 2025 cannot know about curriculum revisions published in March 2025 without additional mechanisms to access current information.
Inconsistency: Ask the same question multiple times, and you may receive different answers – some correct, some incorrect. In our testing, we asked models to translate “phone” into isiZulu and received three different terms across five attempts: “iselula,” “foni,” and “ucingo.” All are technically valid, but consistency in educational terminology is essential.
Bias and Representation: LLMs reflect the biases present in their training data, which can affect how they represent different cultures, historical events, or social issues, particularly problematic for educational content.
The Critical Importance of Human Oversight
These limitations lead to our fundamental principle for AI in educational publishing: AI is a powerful assistant, not an autonomous creator. Every piece of AI-generated content requires human review by subject matter experts who can fact-check, verify alignment with curriculum standards, and ensure pedagogical appropriateness.
We frame this as “human-in-the-loop” workflows, where AI handles the time-consuming tasks of drafting, formatting, and initial content generation, whilst humans provide the critical thinking, expertise, and quality assurance that educational content demands.
Prompting Fundamentals: The C-RAFT Method
If AI systems are to be genuinely useful tools for publishing professionals, users must understand how to communicate effectively with them. Prompt engineering (the practice of crafting effective instructions for AI systems) is not an arcane technical skill but rather a structured communication approach that anyone can learn.
We teach the C-RAFT method, a framework that breaks effective prompts into five essential components:
Context: Providing Situational Background
AI systems lack awareness of your specific situation unless you provide it explicitly. Context includes information about your organisation, the project at hand, the target audience, and any constraints or requirements.
Poor prompt: “Write content about photosynthesis.”
Improved with context: “I am developing Grade 7 Natural Sciences curriculum materials for South African schools, aligned with CAPS. The learners have already covered basic plant structure but have not yet encountered chemical processes.”
The second version gives the AI crucial information about the educational system, curriculum standards, learner age, and prior knowledge—all of which shape appropriate content.
Role: Assigning Expertise
LLMs adjust their output based on the role you assign. Different roles activate different patterns in the model’s training data, producing outputs with appropriate vocabulary, structure, and focus.
Example roles for educational publishing:
- “You are a CAPS-certified Natural Sciences curriculum developer with 15 years of experience creating age-appropriate learning materials.”
- “You are an isiZulu education specialist familiar with PanSALB terminology standards.”
- “You are a literacy specialist who adapts complex texts for learners with reading difficulties.”
The role primes the AI to draw on relevant patterns from its training data, producing more appropriate outputs for your specific use case.
Action: Specifying the Task
The action component clearly states what you want the AI to do. Use specific action verbs and break complex tasks into discrete steps.
Vague action: “Help with this textbook chapter.”
Clear action: “Analyse this Grade 9 Mathematics textbook chapter and create ten multiple-choice assessment questions targeting Bloom’s Taxonomy levels 3-4 (Apply and Analyse). Each question should have four options with one correct answer and three plausible distractors.”
Specificity eliminates ambiguity and produces outputs that require less revision.
Format: Defining Output Structure
AI systems can generate content in virtually any format, but they need explicit instructions about structure, length, and organisation.
Format specifications might include:
- “Provide your response as a markdown table with columns for Question, Option A, Option B, Option C, Option D, and Correct Answer.”
- “Write 300-500 words in three paragraphs, each beginning with a topic sentence.”
- “Generate a JSON object with keys for ‘title’, ‘learning_objectives’, ‘content’, and ‘assessment’.”
Defining format upfront reduces the need for manual reformatting and ensures outputs integrate smoothly into existing workflows.
Tone: Setting the Stylistic Register
The tone component specifies the appropriate voice, formality level, and stylistic characteristics for your content.
Tone specifications might include:
- “Use encouraging, supportive language appropriate for learners who may lack confidence in mathematics.”
- “Write in formal academic style suitable for teacher professional development materials.”
- “Adopt a conversational, accessible tone that explains complex scientific concepts through everyday examples.”
Different educational contexts demand different tones—teacher-facing materials require different language than student-facing content, and primary education materials require different approaches than secondary or tertiary resources.
Practical Application: The Complete Prompt
Here is an example of the C-RAFT method applied to a real publishing task:
Context: “I am developing supplementary literacy materials for Grade 3 isiZulu Home Language learners in rural KwaZulu-Natal schools. Many learners have limited exposure to books outside school. We are working on a unit about community helpers.”
Role: “You are an isiZulu education specialist with expertise in early literacy development and culturally responsive pedagogy.”
Action: “Create a short reading passage about a community nurse that incorporates sight words from the Department of Basic Education’s Grade 3 word list. The passage should include opportunities for phonics practice with the ‘ng’ sound.”
Format: “Provide the passage as a 150-word narrative, followed by five comprehension questions (three literal, two inferential) and a list of vocabulary words with English translations.”
Tone: “Use simple, concrete language with repetitive sentence structures. Include culturally familiar contexts and settings that rural learners will recognise.”
This complete prompt provides the AI with everything it needs to generate appropriate, useful content that requires minimal revision.
Common Pitfalls and How to Avoid Them
Through our training workshops, we have identified several common prompting mistakes:
Assuming the AI knows your context: AI systems do not remember previous conversations unless you are using systems with conversation history. Start each new task with relevant context.
Vague task descriptions: “Make this better” or “improve this content” gives the AI no clear direction. Specify exactly what improvements you need—readability, accuracy, curriculum alignment, etc.
Ignoring output format: Without format specifications, AI will choose its own output structure, which may not match your needs.
Expecting perfection on the first attempt: Effective prompting is often iterative. Use the AI’s first response to identify what needs adjustment, then refine your prompt.
Forgetting constraints: If your content has word limits, reading level requirements, or must avoid certain topics, specify these explicitly.
Working with Source Material: Retrieval-Augmented Generation (RAG)
One of the most powerful techniques for educational publishing involves grounding AI outputs in specific source materials such as curriculum documents, existing textbooks, policy guidelines, and standards frameworks. This approach, known as Retrieval-Augmented Generation (RAG), addresses the hallucination problem by tethering AI outputs to verifiable sources.
What is RAG and Why Does It Matter for Publishing?
Standard LLM interactions rely entirely on the model’s training data—whatever patterns it learned during its initial training process. This creates several problems for educational publishers:
Outdated information: The model knows nothing about curriculum revisions, policy updates, or recent research published after its training cutoff date.
Generic outputs: Without access to your organisation’s specific content guidelines, style standards, or approved terminology, the AI produces generic content that requires extensive revision.
Hallucination risk: When the AI lacks relevant information, it fills gaps with plausible-sounding fabrications rather than admitting knowledge limitations.
RAG solves these problems by allowing the AI to reference specific documents during its response generation. Instead of relying solely on training data, the system retrieves relevant passages from your provided documents and uses these as the foundation for its output.
Types of Source Material for Educational Publishing
Effective RAG implementations in publishing workflows typically involve several categories of source material:
Curriculum Standards and Frameworks: CAPS documents, IEB syllabi, Cambridge Assessment International Education frameworks, or other relevant curriculum specifications. These ensure content aligns with official learning outcomes and assessment standards.
Existing Content Libraries: Previously published textbooks, teacher guides, or supplementary materials. These provide examples of approved content, maintain consistency with existing resources, and prevent duplication of effort.
Terminology Standards: Glossaries, approved terminology lists (such as PanSALB dictionaries for South African languages), or subject-specific vocabulary standards. These ensure consistent use of technical terms across materials.
Style and Editorial Guidelines: Publishing house style guides, accessibility standards, reading level specifications, and editorial policies. These maintain consistency across content creators and align with organisational quality standards.
Assessment Specifications: Examination frameworks, question taxonomies (such as Bloom’s Taxonomy or SOLO), and mark scheme guidance. These ensure assessment materials meet appropriate cognitive demand levels.
Structuring Source Material for Best Results
Not all documents work equally well with RAG systems. We have found several practices that improve AI’s ability to extract and use information from source documents:
Clear section headings: Documents with well-defined sections, headings, and subheadings allow AI systems to locate relevant information more accurately.
Explicit learning objectives: When curriculum documents clearly state learning outcomes, AI can more effectively target content to specific objectives.
Structured formats: Tables, lists, and other structured formats are easier for AI to process than dense paragraphs of prose.
Consistent terminology: Documents using standardised terms throughout reduce ambiguity and improve AI comprehension.
Metadata and tagging: When possible, documents with metadata (grade level, subject, topic tags) help AI systems retrieve the most relevant information.
Practical Example: Generating CAPS-Aligned Content
Consider a common publishing task: creating learning activities aligned with specific CAPS curriculum requirements for Grade 8 Social Sciences. Here is how RAG transforms this workflow:
Traditional approach: A curriculum developer reads through the CAPS document, identifies relevant learning outcomes, manually drafts activities, and cross-references back to the curriculum to verify alignment. This might take 2-3 hours per activity set.
RAG-enabled approach:
- Upload the relevant CAPS document (e.g., Social Sciences Grades 7-9)
- Provide a structured prompt referencing the document:
“Using the attached CAPS Social Sciences document, identify the specific learning outcomes for Grade 8 Term 2 Topic 3 (International Trade). Create five learning activities that:
- Address each of the three specified knowledge areas for this topic
- Progress from lower-order to higher-order cognitive skills
- Include one individual activity, two pair activities, and two group activities
- Provide clear assessment criteria aligned with the CAPS assessment standards for this grade”
The AI retrieves the relevant sections from the CAPS document, extracts the specific learning outcomes and assessment criteria, and generates activities explicitly tied to these requirements. The curriculum developer’s role shifts from drafting to reviewing, verifying accuracy, and refining the generated materials—reducing time to perhaps 30-45 minutes whilst improving curriculum alignment.
When to Use Uploaded Documents vs Web Search
A common question in our workshops is when to rely on uploaded documents versus allowing the AI to search online for information. Our guidance:
Use uploaded documents when:
- You need content aligned with specific curriculum standards or frameworks
- You require consistency with existing published materials
- You are working with proprietary or organisation-specific information
- Accuracy and verifiability are paramount
- You need content based on approved terminology or style guidelines
Use web search when:
- You need current information about recent events, research, or developments
- You are researching emerging topics not covered in existing materials
- You want to identify diverse perspectives or examples
- You are fact-checking or verifying information
Combine both when:
- You need curriculum-aligned content that incorporates current events
- You are developing resources on topics that require both theoretical frameworks (from curriculum documents) and contemporary examples (from online sources)
Many educational publishing tasks benefit from a hybrid approach—using uploaded curriculum documents to ensure alignment and standards compliance, whilst supplementing with web search for current examples, case studies, and real-world applications.
Web Search and Real-Time Information
Whilst RAG allows AI to work with your organisation’s specific documents, web search capabilities enable AI to access current information beyond its training data cutoff. For educational publishing, this opens several valuable use cases whilst also introducing new quality control considerations.
When and Why to Incorporate Online Search
Educational content frequently requires current information:
Current events for Social Sciences: Teaching about international relations, economics, or political systems benefits from recent examples. A Grade 11 Economics lesson on trade agreements should reference actual current agreements, not outdated examples from the AI’s training data.
Scientific developments: Science curricula should reflect recent discoveries and research. Content about space exploration, medical research, or climate science becomes outdated quickly.
Statistical data: Population figures, economic indicators, and research statistics need to be current and properly sourced.
Technology and digital literacy: Educational content about technology, social media, or digital citizenship must reflect current platforms and practices, not conditions from several years ago.
Contemporary literature and media: English and Media Studies curricula often incorporate current books, films, or media that post-date AI training cutoffs.
Ensuring Accuracy and Currency
Whilst web search expands AI capabilities, it also introduces new risks. The AI does not inherently evaluate source credibility—it will retrieve information from academic journals and unreliable websites with equal weight.
Our quality control approach for web-search-enabled content includes:
Explicit source requirements in prompts: Rather than simply asking the AI to search for information, specify source types: “Using recent peer-reviewed research published in academic journals, find three studies on…”
Verification requirements: Instruct the AI to provide URLs for all factual claims: “For each statistic or research finding mentioned, provide the source URL and publication date.”
Multiple source corroboration: Ask the AI to verify key facts across multiple sources: “Confirm this information across at least three authoritative sources before including it in the content.”
Manual fact-checking: Particularly for quantitative claims, dates, and specific events, human reviewers must verify information against primary sources.
Recency filters: When currency is critical, specify date ranges: “Use only sources published within the last 12 months.”
Citation and Attribution Practices
Educational publishing has clear ethical and legal requirements around citation. When using AI-generated content incorporating online sources, several citation considerations arise:
Require citations in initial generation: Build citation requirements into your prompt: “Provide in-text citations in APA 7th format for all referenced sources, with a complete reference list at the end.”
Review and verify citations: AI-generated citations can contain errors—incorrect page numbers, wrong dates, or even hallucinated sources. Every citation requires human verification.
Distinguish between AI synthesis and quoted material: When AI summarises information from sources versus directly quoting, ensure appropriate attribution. We instruct our teams to verify any quoted material against original sources, as AI sometimes generates “quotations” that do not exist in the source.
Maintain audit trails: Keep records of which sources AI consulted during content generation, even if not all appear in final citations. This supports quality assurance and potential future fact-checking.
Practical Example: Current Event Integration
Consider developing a Grade 10 Geography lesson on climate change impacts. The prompt might be:
“Using news articles and scientific reports published in the last 18 months, identify three recent examples of climate change impacts in Africa. For each example:
- Summarise the event or phenomenon
- Explain the environmental and social impacts
- Connect to the CAPS Grade 10 Geography concepts of climate change, environmental sustainability, and human-environment interaction
- Provide the full citation in APA format with URL
Prioritise sources from reputable news organisations (BBC, Reuters, local newspapers) and scientific institutions (CSIR, African Climate and Development Initiative). Verify key facts across multiple sources.”
This structured prompt directs the AI to find current, relevant, geographically appropriate examples whilst specifying source quality standards and citation requirements—producing content that requires verification but significantly reduces research time.
AI for Multilingual South Africa: Translation Workflows
South Africa’s linguistic diversity presents unique challenges for educational publishers who must produce materials in multiple official languages. AI translation offers significant efficiency gains but requires careful structuring to maintain consistency, accuracy, and appropriate terminology.
What Languages Does AI Speak?
Current LLMs have uneven capabilities across South Africa’s 11 official languages:
Strong capability: English, Afrikaans Moderate capability: isiZulu, isiXhosa Limited capability: Sepedi, Setswana, Sesotho, siSwati, Xitsonga, Tshivenda, isiNdebele
This disparity reflects the availability of training data—models trained primarily on English-language internet content naturally perform better with English than with languages less represented online.
However, even with languages of “moderate capability,” we have encountered significant consistency and accuracy issues when using standard translation approaches.
The Problem: Inconsistent Terminology
In early experiments with AI translation, we discovered a critical problem for educational content: terminology inconsistency. We asked multiple AI systems to translate the English word “phone” into isiZulu and received three different results:
- “iselula” (borrowed term from “cellular”)
- “foni” (direct phonetic borrowing from English)
- “ucingo” (traditional isiZulu term meaning “wire” or “cord”)
All three are linguistically valid, but educational materials require consistent terminology across units, textbooks, and assessment materials. A learner encountering “iselula” in one chapter and “ucingo” in another might reasonably wonder if these refer to different concepts.
This problem extends beyond simple vocabulary. Technical terminology, scientific concepts, and curriculum-specific terms require precise, consistent translation—particularly when assessment materials must use identical terminology to content materials.
The Solution: Glossary-Guided Translation
We address this through glossary-guided translation workflows that constrain AI to use approved, standardised terminology. The approach works as follows:
Step 1: Establish authoritative glossaries
We begin with approved terminology sources, primarily PanSALB (Pan South African Language Board) dictionaries, which provide standardised terminology for South African languages. PanSALB terminology undergoes expert linguistic review and represents consensus among language specialists.
For specialised educational content, we supplement PanSALB terms with curriculum-specific glossaries developed by subject matter experts—ensuring terms for mathematical concepts, scientific processes, or technical terminology remain consistent with Department of Basic Education materials.
Step 2: Structure glossaries for AI use
We format glossaries as structured data that AI systems can easily process:
| English | isiZulu | Context |
|———|———|———|
| phone | iselula | PanSALB approved term for mobile phone |
| assessment | ukuhlolwa | Educational context – formal testing |
| photosynthesis | i-photosynthesis | Scientific term – transliterated |
The “Context” column helps the AI apply terminology correctly when English terms have multiple possible translations depending on usage.
Step 3: Constrain translation with explicit glossary instructions
When requesting translation, we provide the glossary and explicitly instruct the AI to use only approved terms:
“Translate the following English text into isiZulu. You MUST use the exact isiZulu terms provided in the attached glossary for any term that appears in the glossary. Do not substitute alternative translations for glossary terms, even if they are linguistically valid. For any term NOT in the glossary, provide a translation and flag it for review.”
Step 4: Flag unknown terms for expert review
Inevitably, content includes terms not yet in the glossary. Rather than allowing the AI to choose translations independently, we instruct it to flag these:
“If you encounter a term not in the provided glossary, translate it but mark it clearly: [NEEDS REVIEW: English term → Proposed isiZulu term]. This flags the term for linguist review.”
This creates a quality control checkpoint where language specialists review AI-proposed translations before they enter the approved glossary.
Workflow Example: CAPS Agricultural Science
In a project translating agricultural science curriculum materials from English to isiZulu, we implemented this approach:
Context: Grade 10 Agricultural Science textbook covering crop production, requiring translation of numerous technical terms (fertiliser, germination, irrigation, etc.)
Process:
- Extracted all technical terms from the English source material (approximately 200 terms)
- Cross-referenced against PanSALB agricultural terminology
- For terms not in PanSALB, consulted the Department of Basic Education’s existing isiZulu Agricultural Science materials
- Created a comprehensive glossary with approved translations
- Used this glossary to constrain all AI translation work
Results:
- 100% consistency in technical terminology across all translated chapters
- Translation time reduced from approximately 8 hours per chapter (human translator) to 2 hours (AI translation + human review)
- Quality remained high as human linguists focused review time on accuracy, natural phrasing, and cultural appropriateness rather than term-by-term translation
Automated Glossary Generation
For organisations building glossaries from scratch, we have developed an AI workflow that accelerates glossary creation whilst maintaining expert oversight:
Step 1: Extract terms requiring translation
An AI agent analyses the English source material and identifies:
- Technical subject-specific terms
- Curriculum-specific terminology
- Terms requiring consistent translation
- High-frequency terms appearing across multiple chapters or units
Step 2: Research approved translations
For each identified term, the AI searches authoritative sources:
- PanSALB terminology databases
- Existing Department of Basic Education materials in the target language
- Academic glossaries from South African universities
- Subject-specific terminology resources
The AI proposes translations with source attribution: “fertiliser → umanyolo (PanSALB Agricultural Terminology, 2018)”
Step 3: Expert review and approval
Linguists and subject matter experts review proposed glossary entries, accepting, modifying, or providing alternative translations. This review ensures:
- Linguistic accuracy and naturalness
- Appropriate register for educational context
- Consistency with broader educational terminology
- Cultural appropriateness
Step 4: Glossary maintenance
As translation work progresses, new terms emerge. The workflow continuously identifies terms not in the current glossary, proposes translations, and routes them to experts for approval—growing the glossary iteratively.
This hybrid approach—AI-assisted glossary generation with expert review—has reduced our glossary development time by approximately 60% whilst maintaining the quality standards required for educational materials.
A Practical Path Forward
The integration of AI into educational publishing represents neither a utopian transformation nor a existential threat, but rather a set of practical capabilities that, when thoughtfully implemented, enhance human expertise rather than replacing it.
Through our workshop programme and ongoing implementation at Cambridge, several core principles have proven essential:
AI as augmentation: The most successful applications position AI as an assistant that handles time-consuming mechanical work, freeing humans for high-value tasks requiring expertise, judgement, and creativity.
Systematic quality control: AI capabilities are only valuable when coupled with rigorous quality assurance—source grounding, fact-checking, expert review, and iterative refinement.
Structured workflows: Moving beyond ad-hoc AI use to systematised workflows ensures consistency, embeds quality control, and scales expertise across organisations.
Human-centric design: Workflows should be designed around human decision-making, with AI providing options and humans making choices.
Continuous learning: Organisations should treat AI implementation as an ongoing learning process, systematically improving prompts, refining workflows, and building collective expertise.
The framework presented in this article – understanding AI capabilities, mastering prompting techniques, implementing RAG, leveraging web search, building multilingual workflows, and systematising processes through automation platforms – provides a replicable approach that other organisations can adapt to their specific contexts.
Educational publishing faces persistent challenges: producing high-quality curriculum materials at scale, serving multilingual populations, maintaining currency in rapidly evolving subjects, and doing all of this within constrained budgets. AI does not solve these challenges automatically, but it does provide powerful capabilities that, when combined with human expertise and systematic quality control, meaningfully enhance what educational publishers can achieve.
Our experience suggests that successful AI adoption requires three parallel tracks: technical implementation (building workflows, training models, integrating systems), skills development (training staff, building expertise, fostering experimentation), and organisational change (addressing anxiety, developing policies, measuring success). Organisations that attend to all three tracks position themselves to capture genuine value from AI capabilities whilst managing risks and maintaining quality standards.
The question for educational publishers is not whether to engage with AI, but how to do so thoughtfully in ways that enhance rather than undermine educational quality, that empower rather than displace professional expertise, and that expand rather than constrain what organisations can achieve in service of learners.