YEEAP · FUSE-AI-101

Foundations of Artificial Intelligence

Reading Material # 4
Module 4: Prompt Engineering and AI–Human
Collaboration

Placid Solutions & Consulting · Florida University Southeast

(FUSE-SPS). InnoLearn Systems ·

YEEAP · FUSE-AI-101
Foundations of Artificial Intelligence
Module 4 Overview: Talking to the Machine — The Art and
Science of Prompt Engineering
Introduction: The Most Practical Module in the Course
Of all six modules in FUSE-AI-101, Module 4 is the one that will most immediately change what
you can do tomorrow. While Modules 1 through 3 built the conceptual and historical foundations
you need to understand AI intelligently, Module 4 puts AI directly in your hands. It teaches you
how to communicate with large language models effectively — how to write inputs that produce
outputs you can actually use, how to design workflows that combine your judgment with AI
capability, and how to evaluate AI outputs with the critical eye of a professional who knows both
the technology's power and its limits.

The skill at the center of this module has a name: prompt engineering. The term sounds technical,
but the concept is straightforward. A prompt is the input you give to an AI model — the text you
type into ChatGPT, Claude, Gemini, or any other large language model. Prompt engineering is the
practice of designing those inputs systematically and skillfully to produce the outputs you need.
It is, in essence, the art of communicating with AI — and like any communication skill, it can be
learned, practiced, and dramatically improved.

This matters enormously for two reasons. First, the quality of what you get from an AI model is
directly, powerfully determined by the quality of what you put in. Two people using the same AI
model for the same task can get radically different results depending on how well they have
designed their prompt. The professional who understands prompt engineering consistently gets
better, more reliable, more useful outputs than the professional who approaches AI as a black box
and hopes for the best. Second, as AI tools become central to professional work across every sector
— healthcare, finance, agriculture, education, government, creative industries — the ability to use
them effectively and critically is becoming a core professional competency, as fundamental as
knowing how to use a spreadsheet or conduct a web search.

Part One: What Large Language Models Actually Do

Before you can communicate with an AI model effectively, you need to understand what it is
actually doing when it responds to you. This understanding is the foundation of effective prompt
engineering — and it is the subject of Sub-Module 4.1.

2
A large language model (LLM) is a neural network trained on vast quantities of text — the contents
of much of the accessible internet, books, scientific papers, code repositories, and other text
sources — with a single training objective: predict the next token in a sequence. A token is
approximately a word or word fragment; the model processes text as a sequence of tokens rather
than characters. Given a sequence of tokens as input, the model produces a probability
distribution over all possible next tokens, samples from this distribution, appends the selected
token to the input, and repeats the process until it generates a complete response.

This process sounds mechanical — and in a technical sense, it is. But the remarkable discovery of
the past decade is that training a very large neural network to predict the next token in text, on a
sufficiently large and diverse corpus, produces a system that has implicitly learned a vast range of
capabilities: factual knowledge, logical reasoning, mathematical problem-solving, code
generation, language translation, summarization, creative writing, and conversation. These
capabilities emerge from the training objective rather than being explicitly programmed — they
are a consequence of what it takes to predict text accurately across the full diversity of human
written expression.

Several technical parameters shape the behavior of LLMs in ways that prompt engineers need to
understand. The context window is the amount of text the model can attend to simultaneously —
its effective working memory for a conversation. Temperature is a parameter that controls the
randomness of the model's sampling: lower temperature produces more deterministic, consistent
outputs; higher temperature produces more varied, creative outputs. The attention mechanism —
the self-attention architecture described in Module 3 — determines how the model relates
different parts of the input to each other when generating each output token. Understanding these
mechanics explains why the same prompt can produce different responses on different runs, why
very long conversations can degrade model performance, and why the model sometimes drifts
from the specific task specified in the prompt.

Conceptual Foundation: The most important conceptual shift in prompt engineering

is moving from thinking about AI as a search engine (type a query, get a result) to
thinking about it as a highly capable but literal collaborator. The model will do what
you tell it — which means if you tell it vaguely or incompletely, you will get vague or
incomplete results. Precision, context, and structure are the core virtues of effective
prompting.

3
Part Two: Core Prompt Engineering Techniques
Prompt engineering encompasses a spectrum of techniques ranging from simple to sophisticated.
At the core are several fundamental approaches that every professional using AI tools should
master. These are the subject of Sub-Module 4.2.

Zero-shot prompting involves giving the model a task with no examples — relying on the model's
pre-trained capabilities to complete the task from your description alone. Most simple AI
interactions are zero-shot: you describe what you want, and the model attempts to produce it.
Zero-shot prompting works well for tasks the model has seen many times in its training data —
writing emails, summarizing text, answering factual questions — but may produce lower quality
results for unusual, specialized, or highly specific tasks.

Few-shot prompting provides the model with one or more examples of the desired input-output
mapping before presenting the actual task. By showing the model what good outputs look like,
few-shot prompting significantly improves performance on tasks where the model needs to
understand a specific style, format, or reasoning pattern. One-shot prompting provides a single
example; few-shot prompting typically provides two to five. The examples serve as implicit
instructions, communicating the task requirements through demonstration rather than
description.

Chain-of-thought prompting instructs the model to reason step by step before producing its final
answer. Rather than asking the model to jump directly to a conclusion, chain-of-thought
prompting asks it to show its reasoning — which both improves accuracy on complex reasoning
tasks and makes the model's logic visible for human review. Simply adding the phrase 'Let's think
step by step' to a prompt can substantially improve performance on multi-step reasoning
problems.

Role and persona assignment instructs the model to adopt a specific identity or professional role
before responding. 'You are an experienced Ethiopian tax attorney...' or 'You are an agricultural
extension worker explaining crop rotation to a smallholder farmer in Amhara...' constrains the
model's response style, vocabulary, knowledge emphasis, and perspective in ways that can
significantly improve relevance for specific use cases. Format control instructions specify the
desired output structure — 'Respond in a numbered list of five items,' 'Format your response as a
table with three columns,' 'Provide your answer in JSON format' — enabling the integration of AI
outputs into downstream systems or workflows.

4
Part Three: Advanced Prompting — System Prompts, RAG, and Agents
Beyond the core techniques, a set of more sophisticated approaches enables AI capabilities that
go significantly beyond what simple conversational prompting can achieve. These are the subject
of Sub-Module 4.3.

System prompts are persistent instructions provided at the beginning of a conversation that
establish the context, role, constraints, and behavior of the AI model for the entire interaction.
Unlike user prompts (which are part of the conversational turn), system prompts are typically
invisible to the end user and are used by developers to configure AI applications for specific use
cases. A customer service chatbot for an Ethiopian telecom company would be configured with a
system prompt that specifies the company's name, the scope of questions the bot should answer,
the tone and language it should use, and the behaviors it should avoid. Understanding system
prompts is essential for building AI-powered applications, not just using AI tools.

Retrieval-Augmented Generation (RAG) is a technique that extends the effective knowledge of an

LLM by connecting it to an external knowledge base. Because LLMs are trained on data up to a
cutoff date and cannot access real-time information, and because they sometimes generate
confident-sounding but incorrect information about topics not well-represented in their training
data, RAG addresses these limitations by retrieving relevant documents from an external database
at query time and including them in the prompt context. An AI assistant for Ethiopian healthcare
workers, for example, could be connected to a database of Ethiopian clinical guidelines and drug
formularies — ensuring that its responses reflect current Ethiopian health system protocols rather
than generic or potentially outdated medical information.

AI agents represent the frontier of prompt engineering — systems where an AI model is given not
just a single task to complete, but a set of tools it can use and a high-level objective it must plan
and execute steps to achieve. An AI agent might be given access to a web search tool, a code
execution environment, a document database, and a calendar system, and asked to research a
topic, synthesize findings, draft a report, and schedule a meeting to present it. Rather than
completing this as a single prompt-response turn, the agent plans a sequence of steps, executes
each one using the available tools, observes the results, and adapts its plan based on what it finds.
Agentic AI is rapidly advancing from research concept to practical deployment, with significant
implications for professional workflows.

5
Part Four: Designing Human-AI Workflows
Prompt engineering is not only about individual interactions with AI — it is about designing the
workflows and processes that determine how AI is integrated into ongoing professional and
organizational activities. This is the subject of Sub-Module 4.4.

Task decomposition is a foundational workflow design principle: breaking complex tasks into
smaller, well-defined subtasks that can be effectively handled by AI, and identifying the points in
the workflow where human judgment, creativity, or accountability is essential and AI should serve
as support rather than decision-maker. A complex report might be decomposed into: AI drafting
individual sections from structured notes, human review and correction of each section, AI
formatting and copyediting, human final review and sign-off.

The question of when to use AI versus when to apply human judgment is one of the most
important professional decisions in the AI era. AI is most valuable when the task is well-defined,
the quality criteria are clear, the volume is large, and the cost of individual errors is manageable.
Human judgment is most important when the task requires contextual understanding that AI
lacks, when the stakes are high enough that errors are unacceptable, when ethical reasoning or
value trade-offs are involved, or when the task requires authentic human relationship and
empathy. Developing a clear personal framework for this judgment — and applying it consistently
— is one of the most important professional skills of the AI era.

Prompt libraries and templates systematize effective prompts for recurring tasks, ensuring that
the best-known prompt formulations are shared across teams and that AI interactions are
consistent, reproducible, and improvable over time. Building and maintaining a personal or
organizational prompt library is an investment in AI productivity that compounds over time —
each new effective prompt formulation becomes a reusable asset.

Part Five: Evaluating and Verifying AI Outputs

Perhaps the most critical professional skill in the entire module is the ability to evaluate AI outputs
critically and verify them against independent sources before relying on them. This is the subject
of Sub-Module 4.5 — and it is the subject on which most AI users, even experienced ones, are
weakest.

The phenomenon of hallucination — in which an LLM generates confident-sounding text that is

factually incorrect — is one of the most important limitations of current AI systems.
Hallucinations are not random noise; they tend to occur on topics where the model has limited

6
training data, where precise factual accuracy is required (specific dates, statistics, citations, legal
provisions, medical dosages), and where the model is interpolating between things it does know
in ways that sound plausible but are wrong. The professional who treats AI output as reliable
without verification is exposed to serious risk — in any context where accuracy matters.

Verification strategies include cross-referencing AI outputs against primary sources (government

websites, peer-reviewed publications, official databases), using search engines to find
corroborating information from multiple independent sources, applying domain expertise to
identify outputs that are implausible or inconsistent with known facts, and asking the model itself
to identify its uncertainty and the basis for its claims (which can help surface potentially
hallucinated content, though this technique has its own limitations). Building personal
verification habits — a consistent practice of checking AI outputs before acting on them — is one
of the most important professional disciplines of the AI era.

Professional Standard: The professional who uses AI well is not the one who uses it
most, or the one who trusts it most. It is the one who has the clearest understanding of
what AI can and cannot do reliably, who applies the right tool to the right task, and
who verifies AI outputs with the same critical rigor they would apply to any other
information source. Effective AI use is an exercise in professional judgment, not a
substitute for it.

Prompt Engineering and African Professional Contexts

For YEEAP learners in Ethiopia and across Africa, prompt engineering has dimensions that are
particularly important to understand. Most large language models have been trained primarily on
English-language text, which means their performance in African languages — including Amharic,
Oromiffa, Tigrinya, Somali, Swahili, Hausa, and hundreds of others — is significantly weaker than
their English performance. Prompting in African languages may produce less accurate, less
coherent, or less culturally appropriate outputs than prompting in English, and learners should
be aware of this limitation when using AI tools in local-language contexts.

At the same time, the ability to prompt effectively in English gives African professionals access to
some of the most powerful AI tools currently available, and the skills developed in this module
translate directly to productive use of those tools for professional tasks — including producing
English-language documents, research, and content at a quality and speed that would otherwise
require far more time and resources. For young professionals entering Ethiopia's growing digital
economy, this capability is a genuine competitive advantage.

7
The workflow design skills in Sub-Module 4.4 are particularly valuable for African organizational
contexts where AI tools are being introduced into workflows that were not designed with AI in
mind. Understanding how to decompose tasks, identify the appropriate role for AI versus human
judgment, and build sustainable AI-augmented work processes is a skill that will be in high
demand across Ethiopian institutions as AI adoption accelerates.

Conclusion: The New Professional Literacy

Prompt engineering — the ability to communicate effectively with AI models, design productive
human-AI workflows, and evaluate AI outputs critically — is rapidly becoming a core professional
literacy, as fundamental in the AI era as word processing and spreadsheet skills were in the digital
era that preceded it. The professional who masters these skills will consistently produce better
results, accomplish more in less time, and navigate the opportunities and risks of AI more
effectively than the professional who does not.

The sub-modules that follow provide the detailed knowledge and practical frameworks you need
to build these skills: the mechanics of LLMs (Sub-Module 4.1), core prompting techniques (Sub-
Module 4.2), advanced prompting including RAG and agents (Sub-Module 4.3), human-AI
workflow design (Sub-Module 4.4), and output evaluation and verification (Sub-Module 4.5).
Engage with each sub-module actively — not just as reading material, but as a framework you test
and apply in your own interactions with AI tools.

Placid Solutions & Consulting · FUSE-SPS · YEEAP Program · April 2026

8
 YEEAP · FUSE-AI-101
Foundations of Artificial Intelligence

Sub-Module 4.1: Understanding Large Language

Models: Foundations and Mechanics

Reading Essay
Tokens, context windows, attention, temperature, and why the same prompt produces
different results

Placid Solutions & Consulting · Florida University Southeast (FUSE-SPS)

InnoLearn Systems · CrossWave Partners
April 2026 — Addis Ababa, Ethiopia

9
Sub-Module 4.1: Understanding Large Language Models
— Foundations and Mechanics
Reading Essay · Estimated Reading Time: 20–24 Minutes

Why Understanding LLM Mechanics Improves Your Prompting

Effective prompt engineering does not require a deep technical understanding of the mathematics
of neural networks. But it does require a working understanding of how large language models
generate their outputs — because the specific mechanics of LLM generation directly explain the
behaviors that users find surprising, frustrating, or exploitable. Why does adding more context to
a prompt improve results? Why does the same prompt sometimes produce wildly different
outputs on different runs? Why do LLMs sometimes give confident answers that are completely
wrong? Why do longer conversations gradually degrade in quality? The answers to all of these
questions are rooted in the mechanics described in this sub-module — and understanding them
transforms you from someone who uses AI tools by trial and error into someone who uses them
with systematic competence.

Tokens: The Units of Language Model Processing

Large language models do not process text as characters or words — they process text as tokens.
A token is approximately a word or a common word fragment; the exact tokenization depends on
the specific tokenizer used by the model. In English, common words like 'the,' 'and,' 'is' are
typically single tokens. Less common words may be split into multiple tokens — 'tokenization'
might become 'token' + 'ization,' for example. Numbers, punctuation, and whitespace also become
tokens. Amharic, Oromiffa, and other Ethiopian languages will be tokenized differently from
English, and because these languages are less represented in training data, they may require more
tokens to represent the same amount of information — a factor that affects both cost and
performance in multilingual AI applications.

Understanding tokens matters for prompt engineering for several reasons. First, most AI APIs
charge for usage based on the number of tokens in the input and output — understanding token
counts helps manage costs. Second, long inputs that approach or exceed the model's context
window will be truncated or handled in ways that can degrade performance. Third, the model's
attention mechanism processes tokens in relation to each other — understanding this helps you
structure prompts so that the most relevant information is positioned advantageously.

10
Context Windows: The Model's Working Memory
Every large language model has a context window — the maximum number of tokens it can
process in a single interaction, encompassing both the input (your prompt) and the output (the
model's response). Early language models had context windows of a few thousand tokens. Current
leading models have context windows ranging from 128,000 tokens (roughly 100,000 words) to
one million tokens or more. This represents a remarkable expansion of what can be included in a
single interaction — an entire book, a large codebase, or a long conversation history can fit within
a modern model's context window.

The context window is effectively the model's working memory for the interaction. Everything the
model knows during the interaction — the conversation history, any documents you have
provided, the instructions you have given — must fit within this window. When the context
window is exceeded, earlier content is typically dropped, which can cause the model to 'forget'
earlier instructions or context.

Practical implications for prompt engineering: front-load your most important instructions.
Research suggests that information at the very beginning and very end of a prompt receives more
attention from the model than information buried in the middle — a phenomenon sometimes
called the 'lost in the middle' problem. For long prompts with complex instructions, repeating the
most critical constraints at the end of the prompt can significantly improve compliance.

Practical Analogy: Think of the context window as a whiteboard on which the entire
conversation is written. The model reads the whole whiteboard before generating each
response. If the whiteboard gets too full and content is erased to make room, the model
loses that context. If important instructions are buried under lots of text in the middle,
they may receive less attention. Keep your most critical instructions prominent — at the
start and end of your prompt.

How LLMs Generate Text: Predicting the Next Token

The generative process of an LLM begins with your input (the prompt) and produces output one
token at a time. At each step, the model processes the entire current context (the prompt plus any
tokens generated so far) through its neural network layers and produces a probability distribution
over all tokens in its vocabulary — assigning a probability to every possible next token. The model
then samples a token from this distribution and appends it to the context. This process repeats
until the model generates a special end-of-sequence token or reaches the maximum output length.

11
 This process has several important implications. First, the model is not looking up answers in a
database — it is generating text probabilistically, token by token. This means there is inherent
variability in outputs, even for identical inputs. Second, because each new token is generated
based on all preceding tokens (including the tokens already generated in the current response),
early choices in the output constrain later ones — the model cannot easily 'go back and revise'
once it has committed to a direction. Third, the model has no access to external information
during generation — everything it knows comes from its training data and from what you have
included in the prompt.

This token-by-token generation process also explains why LLMs sometimes get worse as a
response gets longer. Errors or suboptimal choices early in the generation cascade forward — a
slightly wrong premise in the first paragraph can compound into increasingly wrong conclusions
by the end of a long response. Techniques like chain-of-thought prompting (discussed in Sub-
Module 4.2) address this by structuring the generation process more carefully.

Temperature, Top-p, and Output Variation Controls

When an LLM generates text, it does not simply select the single highest-probability token at each
step — it samples from the probability distribution, introducing controlled randomness into the
generation process. Two parameters control this sampling behavior: temperature and top-p.

Temperature scales the probability distribution before sampling. A temperature of 1.0 means
sampling from the original distribution. Lower temperatures (0.1 to 0.5) make the distribution
more peaked — high-probability tokens become much more likely, producing more deterministic,
consistent outputs. Higher temperatures (1.5 to 2.0) flatten the distribution — lower-probability
tokens become more likely, producing more varied, sometimes more creative but also more
unpredictable outputs.

Top-p (also called nucleus sampling) is an alternative or complementary method for controlling
output variation. Rather than scaling the entire distribution, top-p sampling restricts sampling to
the smallest set of tokens whose cumulative probability exceeds a threshold p. A top-p of 0.9
means sampling only from the tokens that collectively account for 90 percent of the probability
mass, excluding very low-probability tokens regardless of temperature.

Low Temperature (0.1–0.5): Use When... High Temperature (0.8–1.5): Use When...
• You need consistent, reproducible outputs • You want creative variation in outputs

12
• The task has a single correct or best answer • You are brainstorming multiple options or
• You are generating factual summaries or reports approaches

• Reliability matters more than creativity • You are generating creative writing or marketing
copy
• You are processing many similar inputs and
need comparable outputs • You want the model to explore less obvious
possibilities
• You are generating several candidates to select
the best from

Why the Same Prompt Produces DiSerent Results

New users of AI tools are often surprised to discover that the same prompt, submitted to the same
model twice, can produce noticeably different outputs. Understanding why this happens — and
how to manage it — is important for professional use of AI.

The primary source of variation is the stochastic sampling described above. Unless temperature
is set to near-zero, the model will sample differently from the probability distribution each time
it generates a response, producing different token sequences even from identical inputs. This is a
feature, not a bug — it enables the creative variation that makes LLMs useful for brainstorming
and generation tasks. But it means that for tasks where consistency and reproducibility are
important (formal reports, standardized analyses, automated workflows), you should use lower
temperature settings or build verification steps that can identify and filter out unacceptable
outputs.

Additional sources of variation include differences in the model's context (if conversation history
differs between runs), model updates (AI providers regularly update their models, which can
change output behavior), and the non-deterministic nature of the underlying hardware (floating-
point arithmetic on GPUs can produce slightly different results on different runs due to
parallelism and rounding). For mission-critical applications, always test prompts across multiple
runs to understand the distribution of outputs you can expect, not just the result of a single run.

Attention and Why Prompt Structure Matters

The Transformer architecture that underlies modern LLMs processes all tokens in the context
simultaneously using the self-attention mechanism — computing, for each token, a weighted
average of all other tokens in the context, where the weights reflect how relevant each other token

13
is for predicting the current one. This means that every token in your prompt can, in principle,
influence every other token in the model's processing — but in practice, not all tokens receive
equal attention.

Research has documented that LLMs give more weight to information at the very beginning and
very end of long contexts — the 'primacy' and 'recency' effects. Information in the middle of long
contexts receives less attention. This has direct implications for prompt structure: place your most
important instructions at the start of your prompt. If you have long context and need to ensure a
specific instruction is followed, repeat it at the end. Keep critical constraints close to the task
statement, not buried in extensive background material.

The attention mechanism also explains why providing relevant context improves LLM
performance. When you include information in the prompt that is directly relevant to the task —
the specific domain, the intended audience, examples of desired outputs, constraints on format
or content — the model can attend to this information as it generates its response, producing
outputs that are more aligned with your specific needs than a context-free prompt would produce.

Placid Solutions & Consulting · FUSE-SPS · YEEAP Program · April 2026

14
 YEEAP · FUSE-AI-101
Foundations of Artificial Intelligence

Sub-Module 4.2: Core Prompt Engineering

Techniques

Reading Essay
Zero-shot, few-shot, chain-of-thought, role assignment, and format control — the essential
toolkit

Placid Solutions & Consulting · Florida University Southeast (FUSE-SPS)

InnoLearn Systems · CrossWave Partners
April 2026 — Addis Ababa, Ethiopia

15
Sub-Module 4.2: Core Prompt Engineering Techniques
Reading Essay · Estimated Reading Time: 22–26 Minutes

From Theory to Technique: Building Your Prompting Toolkit

Understanding how LLMs work is the foundation. Knowing what to do with that understanding
is the superstructure. This sub-module presents the core techniques of prompt engineering — the
specific practices that consistently improve output quality across a wide range of tasks. These are
not abstract principles; they are actionable methods you can apply immediately in your
interactions with any major AI model.

The techniques presented here form a progression from simple to more structured. Every
professional using AI tools should master all of them. Together, they constitute the core
prompting toolkit — the set of methods that distinguishes a skilled AI user from someone who
types questions and hopes for the best.

Technique 1: Zero-Shot Prompting

Zero-shot prompting is the simplest form of interaction: you give the model a task with no
examples, relying entirely on its pre-trained knowledge and capabilities. Zero-shot prompting is
appropriate for tasks the model handles well based on its training — common writing tasks,
factual questions on well-represented topics, standard translations, basic code generation, and
summarization of provided text.

ZERO-SHOT PROMPT EXAMPLE

Summarize the following paragraph in two sentences, written for a general
audience with no technical background: [Paste your paragraph here]

Zero-shot prompting is most effective when the task is clearly defined, when the model has seen
many examples of similar tasks in its training data, and when you have provided all the context
the model needs to complete the task. Weak zero-shot prompts omit critical context ('Write a
report'), specify the task vaguely ('Make this better'), or leave the output format undefined ('Tell
me about AI'). Strong zero-shot prompts are specific about the task, the audience, the format, the
length, and any constraints.

Specificity Rule: The single most effective improvement to any zero-shot prompt is
adding specificity. Compare 'Write an email' with 'Write a formal 150-word email to a
bank manager in Addis Ababa requesting a meeting to discuss loan options for a small

16
 retail business, using a respectful tone appropriate for Ethiopian professional
correspondence.' The second prompt will produce a dramatically better result.

Technique 2: One-Shot and Few-Shot Prompting

Few-shot prompting improves on zero-shot by providing one or more examples of the desired
input-output mapping before presenting the actual task. The examples communicate the task
requirements through demonstration, which is often more effective than description alone —
particularly for tasks involving specific formats, styles, or reasoning patterns that are difficult to
specify precisely in natural language.

FEW-SHOT PROMPT EXAMPLE

Classify each customer complaint as: BILLING, TECHNICAL, or SERVICE.
Example 1: Customer: "I was charged twice for the same transaction last
month." Category: BILLING Example 2: Customer: "My internet connection
keeps dropping every few hours." Category: TECHNICAL Now classify:
Customer: "The agent I spoke to was rude and unhelpful." Category:

The quality of your examples matters as much as their presence. Use examples that are
representative of the full range of inputs the model will encounter, ensure that the examples are
consistent with each other (if they contradict each other, they will confuse the model), and choose
examples that highlight distinctions the model might find difficult to make without guidance. The
number of examples is a trade-off — more examples improve performance but consume more of
the context window and increase cost. For most tasks, two to five examples strike an effective
balance.

Technique 3: Chain-of-Thought Prompting

Chain-of-thought (CoT) prompting instructs the model to reason step by step before producing
its final answer. Research has demonstrated that this simple modification substantially improves
performance on complex reasoning tasks — mathematical problem solving, multi-step logical
inference, causal analysis, and tasks requiring multiple pieces of information to be synthesized.

The mechanism behind CoT's effectiveness is not fully understood, but the most plausible
explanation is that by externalizing the reasoning process — writing out intermediate steps rather
than jumping to a conclusion — the model is forced to generate each reasoning step as an explicit
token sequence. This makes each step a conditioning context for the next, reducing the probability

17
of reasoning errors that can occur when large inferential gaps must be crossed in a single
generation step.

CHAIN-OF-THOUGHT EXAMPLE — ZERO-SHOT CoT

A smallholder farmer in Oromia has 2 hectares of land. She allocates 60%
to maize and the remaining 40% to teff. If maize yields 3.5 tons per
hectare and teff yields 1.8 tons per hectare, what is her total expected
harvest? Let's think step by step.

CHAIN-OF-THOUGHT EXAMPLE — FEW-SHOT CoT

Q: A clinic has 8 health workers. Each sees 15 patients per day. How many
patients are seen in a 5-day week? A: Let me work through this: 8 workers
× 15 patients = 120 patients per day. 120 patients/day × 5 days = 600
patients per week. The answer is 600 patients. Q: A cooperative has 24
members. Each contributes 500 birr per month. How much does the
cooperative collect in 3 months? A:

Chain-of-thought prompting is particularly valuable for professional tasks that require visible
reasoning: financial analysis, diagnostic reasoning, legal argument construction, policy analysis,
and any task where the reasoning process is as important as the conclusion and where errors in
reasoning can have significant consequences. The visible reasoning chain also makes it easier to
identify where the model went wrong if its conclusion is incorrect — enabling targeted correction
rather than requiring a complete regeneration.

Technique 4: Role and Persona Assignment

Assigning the model a specific role or persona before the task is one of the most reliable ways to
improve output quality for domain-specific tasks. By telling the model who it is and what
perspective it should adopt, you constrain the vocabulary, knowledge emphasis, tone, and framing
it will use in its response.

ROLE ASSIGNMENT EXAMPLES

BASIC: "You are a financial advisor..." STRONGER: "You are an experienced
Ethiopian financial advisor specializing in microfinance for small
businesses in Addis Ababa. You communicate in clear, accessible language,
and you understand the specific regulatory context of the National Bank of
Ethiopia." PROFESSIONAL CONTEXT: "You are a senior agricultural extension
officer from the Oromia Agricultural Development Bureau with 15 years of
field experience working with smallholder teff farmers in the eastern

18
 highlands. Respond to farmer questions as you would in an in-person
consultation."

The more specific and contextually grounded the role assignment, the more effectively it
constrains the model's behavior. A vague role ('You are an expert') provides minimal constraint.
A rich role that specifies the professional context, geographic context, audience, and
communication style provides strong, effective constraint. Role assignment is particularly
powerful for tasks where the same question might receive very different responses depending on
the professional perspective — medical advice, legal guidance, agricultural recommendations, and
policy analysis all benefit from well-constructed role assignments.

Technique 5: Format Control

Specifying the desired output format in your prompt is one of the simplest and most effective
prompt engineering techniques. Without format instructions, models default to a general-
purpose prose response that may not fit your specific use case. With clear format instructions, you
can reliably produce outputs in the exact structure you need.

FORMAT CONTROL EXAMPLES

"Respond in a numbered list of exactly five items." "Format your response
as a table with three columns: Risk, Likelihood (High/Medium/Low),
Mitigation." "Respond in JSON format with the following fields: name,
description, estimated_cost_birr, timeline_weeks." "Write your response in
two paragraphs of no more than 100 words each." "Provide your answer in
the following structure: SUMMARY (1 sentence) KEY FINDINGS (3-5 bullet
points) RECOMMENDATION (1-2 sentences)"

Format control is particularly important when AI outputs will be integrated into downstream
systems or workflows — when the output will be parsed by software, incorporated into a document
template, or used as input to another AI prompt. JSON format control, for example, enables AI
outputs to be directly parsed by applications. Structured output formats
(summary/findings/recommendation) enable AI outputs to be incorporated into standardized
report templates. Consistent formatting also makes AI outputs easier to review, compare, and
quality-control.

19
Combining Techniques: The CRRF Framework
The most effective prompts typically combine multiple techniques. A useful framework for
structuring complex prompts is the CRRF framework: Context, Role, Request, Format. Context
provides the background information and situational constraints the model needs. Role assigns
the professional identity and perspective. Request specifies the task clearly and precisely. Format
specifies the desired output structure.

CRRF FRAMEWORK EXAMPLE

CONTEXT: "Ethiopia's National Bank has recently introduced new micro-
lending regulations that require mobile money lenders to verify borrower
identity through the Fayda national ID system before disbursing loans
above 5,000 birr." ROLE: "You are a compliance officer at a mobile money
lending company in Addis Ababa." REQUEST: "Write a brief internal memo to
the lending operations team explaining the new requirement and the three
specific steps they need to take to implement it." FORMAT: "The memo
should be no longer than 250 words, use formal Ethiopian business memo
format, and include a numbered list of the three implementation steps."

The CRRF framework is not a rigid prescription — not every prompt needs all four components.
But using it as a checklist when you are dissatisfied with AI output quality can help you diagnose
what is missing. Vague outputs often reflect insufficient context or an underspecified request.
Inappropriate tone or perspective reflects a missing or weak role assignment. Difficult-to-use
outputs reflect missing format instructions.

Placid Solutions & Consulting · FUSE-SPS · YEEAP Program · April 2026

20
 YEEAP · FUSE-AI-101
Foundations of Artificial Intelligence

Sub-Module 4.3: Advanced Prompting: System

Prompts, RAG, and Agents

Reading Essay
Configuring AI behavior, grounding responses in real data, and building AI systems that plan
and act

Placid Solutions & Consulting · Florida University Southeast (FUSE-SPS)

InnoLearn Systems · CrossWave Partners
April 2026 — Addis Ababa, Ethiopia

21
Sub-Module 4.3: Advanced Prompting — System
Prompts, RAG, and Agents
Reading Essay · Estimated Reading Time: 20–24 Minutes

Beyond Conversational Prompting: Building AI Systems

The techniques in Sub-Module 4.2 — zero-shot, few-shot, chain-of-thought, role assignment,
format control — are primarily tools for individual interactions: a human types a prompt, the
model produces a response. The techniques in this sub-module go further. They are tools for
configuring AI systems that operate across many interactions, that connect AI models to external
knowledge and capabilities, and that enable AI to plan and execute multi-step tasks
autonomously. These techniques sit at the boundary between prompt engineering and AI
application development — understanding them is essential for anyone who will build,
commission, or govern AI-powered applications, not just use pre-built AI tools.

System Prompts: Configuring AI Behavior at Scale

When you interact with an AI application — a customer service chatbot, a document analysis tool,
an AI writing assistant — the behavior you observe is shaped not only by the messages you type
but by a set of persistent instructions that are provided to the model before your conversation
begins. These are called system prompts.

A system prompt is a special message — typically invisible to end users — that establishes the AI's
identity, scope, constraints, and behavior for the entire interaction. System prompts are used by
application developers to configure general-purpose AI models (like GPT-4 or Claude) for specific
use cases. They specify what the AI's role is, what topics it should address and which it should
decline, what tone and language it should use, what format it should produce its outputs in, and
what safety and compliance constraints it must observe.

EXAMPLE SYSTEM PROMPT — Ethiopian Healthcare Chatbot

You are a health information assistant for the Ethiopian Federal Ministry
of Health. Your role is to provide accurate, evidence-based health
information to Ethiopian citizens. SCOPE: You answer questions about
common illnesses, preventive health, nutrition, maternal and child health,
and when to seek medical care. You do not diagnose diseases or prescribe
medications. ESCALATION: For any question about symptoms that may
indicate a serious condition, always advise the user to visit their
nearest health center or call 907 (Ethiopia's health emergency line).

22
 LANGUAGE: Respond in the same language the user writes in (Amharic,
Oromiffa, or English). Use simple, accessible language appropriate for
adults with primary school education. TONE: Warm, reassuring, and
respectful. Use culturally appropriate terms.

Understanding system prompts matters for AI practitioners for several reasons. First, if you are
building AI-powered applications, designing effective system prompts is one of the most
important skills you need — a well-designed system prompt can transform a general-purpose LLM
into a highly specialized, reliable, and safe application for a specific context. Second,
understanding that AI assistants are shaped by system prompts helps you interpret their behavior
— an AI that refuses certain requests or consistently responds in a particular way is likely
following system prompt instructions, not expressing an inherent limitation of the underlying
model. Third, system prompts themselves can have vulnerabilities — prompt injection attacks,
where malicious users attempt to override system prompt instructions through their inputs, are a
real security concern for AI applications.

Retrieval-Augmented Generation: Grounding AI in Real Data

One of the most significant limitations of language models is that their knowledge is frozen at
their training cutoff date. An LLM cannot tell you the current price of teff in Addis Ababa markets,
the most recently published Ethiopian Central Statistical Agency population estimate, or the
contents of a regulation published last month. It also has limited and potentially inaccurate
knowledge about specialized topics that were underrepresented in its training data — including
many topics specific to Ethiopia and Africa.

Retrieval-Augmented Generation (RAG) addresses this limitation by connecting the LLM to an

external knowledge base that can be queried at the time of each user interaction. When a user
submits a query, the RAG system first retrieves relevant documents or passages from the
knowledge base (using semantic search — finding documents that are meaningfully similar to the
query, not just keyword-matching), includes those retrieved passages in the prompt context, and
then passes the augmented prompt to the LLM. The LLM generates its response based on both its
pre-trained knowledge and the retrieved passages, which ground the response in current, domain-
specific information.

RAG ARCHITECTURE — How It Works

USER QUERY: "What are the current NBE guidelines for mobile lending
interest rates?" STEP 1 — RETRIEVAL: System searches knowledge base of

23
 NBE directives and regulations, retrieves the most relevant passages from
the current mobile lending directive. STEP 2 — AUGMENTED PROMPT TO LLM:
"Based ONLY on the following official NBE directive text: [retrieved
passages], answer the user's question about current interest rate
guidelines. If the answer is not contained in the provided text, say so."
STEP 3 — GROUNDED RESPONSE: LLM generates a response citing specific
provisions from the actual current directive, rather than potentially
outdated or hallucinated general knowledge.

RAG systems have significant practical applications across African professional contexts. An AI
legal research assistant for Ethiopian lawyers, grounded in a database of Ethiopian legislation and
court decisions, can provide more accurate and current legal guidance than a generic LLM. An
agricultural advisory system for Ethiopian extension workers, grounded in a database of locally
validated agronomic research and seasonal weather data, can provide better-calibrated
recommendations than a system relying on global training data alone. A medical decision support
system grounded in the Ethiopian standard treatment guidelines can ensure that AI-generated
clinical guidance reflects the Ethiopian therapeutic context rather than potentially inappropriate
international protocols.

Building a RAG system requires several components beyond the LLM itself: a knowledge base (a
structured collection of documents, regularly updated), a document embedding model (which
converts documents and queries into numerical vectors for semantic search), a vector database
(which stores the embeddings and enables fast similarity search), and an orchestration layer
(which coordinates the retrieval and generation steps). Open-source tools including LangChain,
LlamaIndex, and various vector databases (Pinecone, Weaviate, Chroma) have made building
RAG systems significantly more accessible in recent years.

AI Agents: Planning, Tool Use, and Multi-Step Execution

The most sophisticated frontier of prompt engineering is agentic AI — systems where a language
model is given not just a task to complete in a single response, but a set of tools it can use and a
high-level objective it must plan and execute multiple steps to achieve. AI agents represent a
qualitative shift in AI capability: from a system that responds to individual prompts to a system
that can autonomously plan and carry out complex, multi-step tasks.

The conceptual core of an AI agent is the ReAct loop — a cycle of Reasoning, Acting, and
Observing. Given an objective, the agent reasons about what steps are needed to achieve it, selects
an action from its available tools (which might include web search, code execution, database

24
query, file operations, API calls, or interactions with other AI models), executes the action,
observes the result, and then reasons about what to do next based on what it has observed. This
loop continues until the agent determines that the objective has been achieved or that it cannot
proceed.

AGENT EXAMPLE — Research and Report Generation

OBJECTIVE: "Research recent developments in AI applications for
smallholder agriculture in East Africa, and produce a 500-word summary
report." AGENT REASONING LOOP: Step 1 → REASON: I need to search for
recent academic and news sources on this topic. ACT:
[web_search("AI agriculture smallholder East Africa 2024-2025")]
OBSERVE: [Search results returned with 8 relevant articles] Step 2 →
REASON: I should read the most relevant articles in detail. ACT:
[fetch_url("[Link] OBSERVE: [Article content
retrieved] Step 3 → REASON: I now have enough information to draft the
report. ACT: [generate_report(sources, outline)]
OBSERVE: [Draft report produced] Step 4 → REASON: Report complete.
Objective achieved. FINAL RESPONSE: [Delivers formatted report to
user]

AI agents are beginning to appear in production applications across multiple domains. GitHub
Copilot Workspace can plan and implement multi-file code changes from a high-level task
description. AI research assistants can autonomously search the web, read papers, and synthesize
findings into reports. AI customer service agents can query customer databases, look up account
information, and take actions on behalf of customers without human intervention.

The governance challenges raised by AI agents are significant. When an AI agent takes
autonomous actions — sending emails, executing code, modifying files, making API calls — the
consequences can be difficult to reverse and can occur before humans have an opportunity to
review them. Careful design of agent permissions (what tools the agent can access), action scope
(what actions it is permitted to take), and human-in-the-loop checkpoints (what decisions require
human approval before the agent can proceed) is essential for safe agentic AI deployment.

Multi-Turn Conversation Design and Memory Management

Most professional AI use occurs not in single-prompt interactions but in extended, multi-turn
conversations — dialogues where context accumulates across multiple exchanges. Effective multi-
turn conversation design is an important prompt engineering skill that goes beyond crafting
individual prompts.

25
Memory management in multi-turn conversations involves ensuring that the model retains
relevant context from earlier in the conversation without consuming the entire context window
on conversation history. Common approaches include: summarization (periodically summarizing
earlier conversation into a compact form that preserves key information while reducing token
count), explicit re-statement (restating the key context and constraints at relevant points in the
conversation rather than relying on the model to retain them from early exchanges), and
structured context management (maintaining a structured summary of key facts, decisions, and
constraints established in the conversation, which is prepended to each turn).

For professional AI workflows involving long, complex tasks — drafting a detailed proposal,
conducting a multi-stage analysis, developing a curriculum or training program — multi-turn
conversation design is what separates a productive AI-assisted workflow from a frustrating series
of disconnected interactions. Invest time in thinking about how to structure the conversation:
what to establish upfront, what to introduce progressively, and how to manage context as the
conversation extends.

Placid Solutions & Consulting · FUSE-SPS · YEEAP Program · April 2026

26
 YEEAP · FUSE-AI-101
Foundations of Artificial Intelligence

Sub-Module 4.4: Designing Human-AI Workflows

Reading Essay
Task decomposition, knowing when to use AI vs. human judgment, prompt libraries, and
collaboration design

Placid Solutions & Consulting · Florida University Southeast (FUSE-SPS)

InnoLearn Systems · CrossWave Partners
April 2026 — Addis Ababa, Ethiopia

27
Sub-Module 4.4: Designing Human-AI Workflows
Reading Essay · Estimated Reading Time: 20–24 Minutes

From Individual Prompts to Productive Systems

Individual prompt engineering skill — the ability to craft effective prompts for specific tasks — is
necessary but not sufficient for realizing the full professional value of AI. The larger opportunity
is designing entire workflows that systematically integrate AI capabilities with human judgment,
institutional knowledge, and professional accountability. This sub-module examines the
principles and practices of human-AI workflow design: how to decompose complex tasks, how to
identify where AI adds most value and where human judgment is essential, how to build and
maintain prompt libraries, and how to design AI-assisted workflows for common professional
activities.

This shift from individual interaction to workflow design represents a maturation in how AI is
used — from experimenting with AI to building AI-augmented professional capability.
Organizations that invest in workflow design alongside individual AI skill development
consistently get dramatically better returns on their AI investments than those that focus only on
tool access.

Task Decomposition: Breaking Complex Work into AI-Tractable Units

Task decomposition is the foundational skill of human-AI workflow design. Complex professional
tasks — writing a comprehensive report, conducting a thorough market analysis, designing a
training curriculum, building a project proposal — are rarely well-handled by a single AI prompt.
They involve multiple distinct subtasks, each with different information requirements, different
quality criteria, and different roles for AI assistance versus human judgment.

Effective task decomposition for AI-augmented workflows involves: identifying the discrete
subtasks within a complex task, assessing the role of AI in each subtask (AI-primary, AI-assisted,
or human-primary), identifying the information and context each subtask requires and where that
information comes from, and sequencing the subtasks in an order that manages dependencies
and information flow effectively.

TASK DECOMPOSITION EXAMPLE — Developing a Training Program

COMPLEX TASK: "Develop a two-day leadership training program for mid-level
managers at a bank." DECOMPOSED SUBTASKS: 1. [Human-primary] Define
learning objectives, audience profile, and organizational context 2. [AI-

28
 assisted] Research leadership development best practices and benchmark
programs 3. [AI-primary] Generate draft curriculum structure with module
titles and time allocations 4. [Human review] Review and refine curriculum
structure — add organizational context, adjust priorities 5. [AI-primary]
Draft detailed content outlines for each module 6. [Human-primary] Review
for accuracy, relevance, and alignment with organizational culture 7. [AI-
assisted] Draft facilitator guides and participant handouts from approved
outlines 8. [Human review] Final review, approval, and sign-off before
delivery KEY INSIGHT: AI handles research, drafting, and structuring.
Humans own objectives, context, quality assurance, and accountability.

The discipline of task decomposition also surfaces an important insight: many complex
professional tasks have a natural structure of alternating AI and human contributions. AI excels
at the high-volume, structurally defined work of drafting, organizing, researching, and formatting.
Humans excel at the judgment-intensive work of defining objectives, providing organizational
context, evaluating quality and relevance, and making decisions with accountability. Designing
workflows that explicitly alternate between these two modes — rather than trying to do everything
in a single AI interaction — produces dramatically better results.

When to Use AI vs. When to Apply Human Judgment

One of the most important professional skills in the AI era is knowing when AI is the right tool for
a task and when human judgment is essential. This is not a binary decision — it is a spectrum, and
most complex tasks involve AI and human contributions in different proportions at different
stages. The following framework provides a starting point for making these decisions
systematically.

AI Is Most Valuable When... Human Judgment Is Essential When...

• The task is well-defined with clear quality criteria• High-stakes decisions with significant
• Volume is large and consistency matters consequences for individuals

• The task involves searching, structuring, or • Ethical trade-oFs and value judgments are
drafting from a large information space involved

• The content domain is well-represented in AI • Deep contextual knowledge of the specific

training data organization or community is required

• Speed matters more than perfect quality • Authentic personal or professional relationship
and empathy matter

29
• The task is tedious and low-stakes for individual • Legal, medical, or professional accountability is
errors at stake
• The output will be reviewed and refined by a • Creative originality and genuine novelty are
human before use required
• The domain is underrepresented in AI training
data (e.g. local contexts)

For Ethiopian professional contexts, a particularly important criterion is local knowledge depth.
AI models have limited, often inaccurate knowledge of Ethiopian-specific institutional contexts,
regulatory frameworks, cultural norms, and community dynamics. Tasks that require deep
knowledge of how a specific Ethiopian institution works, how a specific regulation is interpreted
in practice, or how a specific community would respond to a proposed intervention require human
judgment — AI can provide structural scaffolding and draft language, but the locally-grounded
content must come from human expertise.

Professional Heuristic: A useful heuristic for the AI era: use AI for the work you would
delegate to a competent, knowledgeable intern who lacks deep organizational context
and whose work you will always review before using. Retain direct human judgment
for the work that requires deep organizational context, accountability, ethical
reasoning, or authentic relationship. The distinction is not about AI capability in
general — it is about what AI can reliably do in your specific context with your specific
data.

Building and Maintaining Prompt Libraries

A prompt library is a structured collection of effective, tested prompts for recurring tasks — an
organizational asset that codifies AI prompting knowledge and makes it reusable, shareable, and
improvable over time. Building a prompt library is one of the highest-return investments an
individual professional or organization can make in AI productivity.

A well-organized prompt library for a professional context typically includes: prompt templates
(with placeholders for task-specific content), notes on the intended use case and quality criteria
for each template, examples of inputs and outputs from successful uses, version history (tracking
improvements made over time), and tags or categories for easy retrieval. Prompt libraries should
be treated as living documents — updated as new techniques are discovered, as model behavior
changes with model updates, and as the team learns what works best for specific tasks.

30
For organizational prompt libraries, governance is important: who is responsible for maintaining
and updating the library, how are new prompts vetted before being added, and how are outdated
prompts retired when model updates change their performance? These are the same questions
that apply to any organizational knowledge asset, and the same answers apply: clear ownership,
defined review processes, and regular maintenance cycles.

PROMPT LIBRARY ENTRY EXAMPLE

PROMPT ID: COMM-EMAIL-FORMAL-001 TASK: Formal email — external stakeholder
communication USE CASE: Drafting formal emails to government officials,
partner organizations, or clients TEMPLATE: "You are a professional
communications officer at [ORGANIZATION_NAME] in Addis Ababa, Ethiopia.
Write a formal email in [LANGUAGE] to [RECIPIENT_TITLE/NAME] at
[RECIPIENT_ORGANIZATION]. Purpose: [STATE PURPOSE IN ONE SENTENCE] Key
information to convey: [LIST 3-5 KEY POINTS] Desired response/action:
[WHAT DO YOU WANT THE RECIPIENT TO DO?] Tone: [Formal/Semi-
formal/Respectful-but-direct] Length: No more than [WORD_COUNT] words.
Use appropriate Ethiopian business correspondence conventions." NOTES:
Best for outbound communications to formal institutional recipients.
Verify regulatory details independently before sending. Tested with Claude
3.5 and GPT-4o. LAST UPDATED: March 2026

AI-Assisted Workflows for Common Professional Tasks

The following outlines AI-assisted workflow patterns for four common professional activities:
writing and editing, research and synthesis, data analysis, and project planning. Each workflow is
designed to maximize AI contribution while preserving essential human oversight.

Writing and Editing Workflows

The most effective AI-assisted writing workflows do not start with 'Write a report about X.' They
start with human-generated structure: an outline, key arguments, the specific evidence or data to
be included, and the audience and purpose. AI then drafts content section by section from these
human-provided anchors, the human reviews and refines each section, and AI assists with
copyediting, formatting, and consistency checking in a final pass. This produces writing that is
structurally and substantively controlled by the human author, with AI providing the drafting
velocity that makes the process sustainable at scale.

31
Research and Synthesis Workflows
AI research workflows typically follow a two-phase structure: a discovery phase (using AI to
identify relevant information, summarize sources, and map the information landscape) and a
synthesis phase (using AI to help organize findings into a coherent narrative, while humans apply
judgment about which sources are most credible and relevant). The critical human contribution
in research workflows is source verification — AI-identified sources and AI-generated summaries
must always be checked against the original sources before being cited or relied upon.

Data Analysis Workflows

AI-assisted data analysis typically involves AI generating and executing code (in Python or R) to
perform specified analyses, the human reviewing the code for logic errors and the outputs for
plausibility, and AI generating narrative interpretations of the results. The key human
contribution is the analytical judgment: deciding which analyses to run, interpreting results in
organizational context, and translating statistical findings into actionable conclusions. AI
dramatically accelerates the execution of analysis; it does not replace the intellectual judgment
that determines what analysis is worth running.

Placid Solutions & Consulting · FUSE-SPS · YEEAP Program · April 2026

32
 YEEAP · FUSE-AI-101
Foundations of Artificial Intelligence

Sub-Module 4.5: Evaluating and Verifying AI

Outputs

Reading Essay
Hallucinations, fact-checking, bias recognition, and building professional verification habits

Placid Solutions & Consulting · Florida University Southeast (FUSE-SPS)

InnoLearn Systems · CrossWave Partners
April 2026 — Addis Ababa, Ethiopia

33
Sub-Module 4.5: Evaluating and Verifying AI Outputs
Reading Essay · Estimated Reading Time: 20–24 Minutes

The Critical Competency No One Talks Enough About

Of all the skills in Module 4, the one that is most underemphasized in popular AI discourse — and
most essential for professional practice — is the ability to evaluate AI outputs critically before
acting on them. Most AI marketing focuses on what AI can do: the reports it can draft, the code it
can write, the answers it can provide. Much less attention is paid to what AI gets wrong, how often
it gets it wrong, and how professionals should handle AI outputs to protect themselves, their
organizations, and the people they serve from the consequences of AI errors.

This sub-module fills that gap. It examines the phenomenon of hallucination — the most
consequential failure mode of current AI systems — in detail, provides concrete strategies for fact-
checking and source verification, addresses the problem of bias in AI outputs, and presents a
practical framework for building personal verification habits that can be applied consistently
across professional contexts.

Hallucinations: Understanding AI's Most Dangerous Failure Mode

Hallucination refers to the phenomenon in which a large language model generates confident-
sounding text that is factually incorrect. The term is borrowed from psychology, where a
hallucination is a perception without a corresponding external stimulus — something that seems
real but is not. In AI, hallucination is a generation without a corresponding factual basis — an
output that sounds authoritative but is wrong.

Hallucinations are not random errors. They have characteristic patterns that, once understood,
become more predictable and manageable. The following are the most common hallucination-
prone contexts:

• Specific quantitative facts — statistics, percentages, dates, prices, population

figures. The model may have seen many statistics in its training data but produce a
plausible-sounding number that does not correspond to any specific source.
• Citations and references — specific paper titles, journal names, authors, and
publication years. LLMs notoriously hallucinate citations that appear credible but do
not exist, often combining real elements (a real author, a real journal, a real research
area) in fabricated combinations.

34
 • Specific legal and regulatory provisions — the exact text of laws, regulations, court
decisions, or contractual clauses. The model may accurately convey the general
intent of a legal provision while fabricating specific section numbers, thresholds, or
conditions.
• Recent events and developments — anything that occurred after the model's training
cutoH date. The model has no information about post-cutoH events and may
generate plausible-sounding but entirely fabricated claims about recent
developments.
• Domain-specific technical details — precise medical dosages, engineering
specifications, financial calculations, agricultural yield figures specific to a particular
region or crop variety. The more specific the technical detail, the higher the
hallucination risk.
• Information about specific people, organizations, and local contexts — particularly
for individuals, organizations, and places that were underrepresented in the training
data.

Why This Is Dangerous: The insidious quality of hallucinations is that they are
generated with the same confident, fluent prose as correct information. There is
typically no linguistic signal that distinguishes a hallucinated fact from a correct one in
the model's output. This is why external verification is not optional — it is a
professional obligation whenever AI-generated content will be used for any
consequential purpose.

Why Hallucinations Occur

Understanding why hallucinations occur helps explain when they are most likely and how to
reduce their frequency. LLMs generate text by predicting the most likely next token given the
preceding context. When the model encounters a prompt that calls for specific factual information
it has not encountered frequently in its training data, it cannot simply 'look up' the answer — it
can only generate the most plausible-seeming continuation of the text. In many cases, the most
plausible continuation is a specific-sounding fact rather than an admission of uncertainty —
because in the text the model was trained on, specific facts appeared far more frequently than
admissions of ignorance.

The model's training objective — predicting the next token accurately across a vast corpus of text
— incentivizes fluency and plausibility, not factual precision. A model that frequently said 'I don't
know the exact figure' would score worse on next-token prediction tasks than one that confidently
generated plausible-sounding specific figures. This is a fundamental limitation of the current

35
generation of LLMs that cannot be fully resolved through prompting — only through external
verification and appropriate system design.

Techniques that can reduce (though not eliminate) hallucination frequency include: asking the
model to explicitly state its uncertainty and the basis for its claims, instructing the model to say 'I
don't know' when it lacks specific information, using RAG to ground the model's responses in a
verified knowledge base (as described in Sub-Module 4.3), and using chain-of-thought prompting
to surface the model's reasoning, which makes it easier to identify potentially hallucinated claims.

Fact-Checking and Source Verification Strategies

The professional standard for AI-assisted work is: verify before you rely. Every AI-generated claim
that will be used for a consequential purpose — included in a report, cited in a proposal,
communicated to clients or stakeholders, used to inform a decision — should be verified against
an independent source. The following verification strategies form a practical toolkit.

Primary Source Verification

Find the original source for any specific fact, statistic, legal provision, or citation. For
statistics, go to the original dataset or publication. For legal provisions, go to the official
1 legislative text. For citations, look up the paper directly.
Example: AI claims 'Ethiopian inflation was 28% in Q3 2023' → Verify
at [Link] or IMF data portal

Multi-Source Cross-Reference
For important factual claims, find at least two independent sources that corroborate the
claim. If multiple credible sources agree, confidence is higher. If sources disagree, the
2 discrepancy itself is important information.
Example: AI describes a regulatory requirement → Check both the
regulation itself and a compliance guide from a law firm

Domain Expert Consultation

For highly specialized technical claims — medical, legal, engineering, agricultural — have
the AI output reviewed by a qualified domain expert before relying on it for consequential
3 decisions.
Example: AI drafts medical protocol recommendations → Have reviewed by
a qualified physician before clinical use

Recency Check
For any claim about current conditions, current laws, current prices, or recent events, verify
4 that the source is current. AI training data has a cutoff date; information that was accurate
at training time may be outdated.

36
 Example: AI describes current mobile lending regulations → Check the
NBE website for the most recent directive date

Self-Consistency Check
Ask the AI to explain the basis for specific claims and to identify what it is uncertain about.
While this does not guarantee accuracy, it can surface claims where the model is generating
5 from sparse training data rather than from frequently-confirmed facts.
Example: Follow-up prompt: 'What is the source for the statistic you
cited? How confident are you in that figure?'

Recognizing and Mitigating Bias in AI Outputs

Beyond hallucination, AI outputs can be systematically biased in ways that reflect biases in the
training data, the choices made in model development, and the social contexts in which the model
was trained. Understanding how to recognize and respond to bias in AI outputs is an essential
professional skill.

Several types of bias commonly appear in LLM outputs. Representation bias occurs when the
model's training data underrepresents certain groups, perspectives, or contexts — producing
outputs that systematically reflect the perspectives of over-represented groups. An LLM asked to
describe 'a typical professional' may default to describing someone who does not reflect the
diversity of Ethiopian or African professional contexts. An LLM asked for examples may default
to examples from North American or European contexts when African examples are equally or
more relevant.

Cultural bias occurs when the model's outputs reflect cultural assumptions, values, and norms
that are specific to the cultural contexts dominant in its training data — particularly Western,
English-speaking contexts. This can manifest in advice that assumes cultural contexts (legal
systems, family structures, business practices, communication norms) that do not apply in
Ethiopian or African settings. Language bias occurs when the model's performance is significantly
better in English than in other languages, producing outputs of inconsistent quality when
prompting in local languages.

Practical strategies for managing bias in AI outputs include: explicitly specifying the cultural,
geographic, and institutional context in your prompts (giving the model the contextual
information it needs to produce locally appropriate outputs), critically reviewing AI outputs for
assumptions that may not apply in your context (particularly legal, economic, or cultural
assumptions), using local experts to review AI-generated content for cultural appropriateness,

37
and maintaining awareness that AI tools are more reliable for some contexts and purposes than
others.

Building Personal Verification Habits and Quality Protocols

Systematic verification is not a one-time action — it is a professional habit that must be built and
maintained over time. The following quality protocol provides a practical framework for
developing consistent AI output verification practices.

Before using any AI-generated content for a consequential purpose, ask: Is there a specific factual
claim here that could be wrong? Is there a citation or reference that I have not verified? Is there a
legal or regulatory provision that needs to be checked against the official source? Is there a
calculation or numerical result that I should verify independently? Is there an assumption about
context, culture, or institutional practice that may not apply to my specific situation? Is this
content recent enough, or could it reflect outdated information?

For high-stakes outputs — formal reports, proposals submitted to funders or clients, regulatory
submissions, medical or legal guidance, policy recommendations — build a structured review step
into your workflow where these questions are addressed systematically before the content is used.
Treat AI-generated content as a first draft that requires professional review, not as a final product
that can be used directly.

Professional Accountability Standard: The standard for using AI professionally is

not 'Is this output probably correct?' It is 'Have I verified the claims in this output that I
will be accountable for?' The professional who applies this standard consistently builds
a reputation for reliability and accuracy. The professional who skips verification
because the AI output looks plausible exposes themselves to the professional
consequences of AI errors that they could have caught.

Building verification habits is easier when it is systematized into workflows rather than relying on
memory and discipline alone. Create checklists, build review steps into document templates, use
commenting tools to flag AI-generated sections that need verification, and cultivate the habit of
being explicit in documents about which content is AI-assisted and which has been independently
verified. These practices protect you professionally, produce more reliable work, and help build
organizational capability in responsible AI use.

Placid Solutions & Consulting · FUSE-SPS · YEEAP Program · April 2026

38
References
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of
stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on
Fairness, Accountability, and Transparency, 610–623.
[Link]

Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Chen, S., Chen, Y.,
Chi, E. H., Chowdhery, A., Demszky, D., Farhadi, A., Fei-Fei, L., Ganguli, D., Grossman, S.,
Guestrin, C., Gururangan, S., Ho, D., Icard, T., … Liang, P. (2021). On the opportunities and risks
of foundation models. arXiv. [Link]

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam,
P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R.,
Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., … Amodei, D. (2020). Language models are few-
shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.

OpenAI. (2023). GPT-4 technical report. arXiv. [Link]

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., &
Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing
Systems, 30.

39