AI Application and Prompt Engineering

Course Description

AI Application and Prompt Engineering prepares students to communicate with and leverage the capabilities of AI language models, exploring the craft of constructing effective prompts that yield accurate, relevant, and creative outputs from AI systems.

Within the SCNS taxonomy, CAI is the Computing and Artificial Intelligence prefix — a recent addition to Florida's course inventory. Daytona State publishes the non-suffixed CAI1320 at 3 credits, prerequisite COP1000, offered spring. The C suffix denotes a combined lecture-and-laboratory format, giving approximately 60 contact hours at the DSC computing convention of 20 hours per credit for C-suffixed courses.

A course on this subject in a college catalog is necessarily provisional, and it is worth being honest about that. The tools change every few months; the underlying competencies do not. The durable content here is not any particular prompting trick — it is understanding what these systems are, what they cannot do, how to verify their output, and how to integrate them into real work responsibly.

⚠ Suffix and catalog-year note

The Florida course inventory carries CAI1320C; Daytona State publishes CAI1320 — same title, no suffix. The suffix is part of the course number and SCNS equivalency does not cross it. This repository's audit found the pattern in 246 of 1,889 queue identifiers.

Note also that this course appears in DSC's 2025–2026 catalog and not in the 2024–2025 edition — the CAI prefix and its programme are new. Check the current catalog year when confirming availability, prerequisites, or credit value for anything in this prefix.

Learning Outcomes

Required Outcomes

• Describe what a large language model is and, at a conceptual level, how it produces output.
• Describe the capabilities and the inherent limitations of current generative AI systems.
• Explain why these systems produce confident, fluent output that may be factually wrong.
• Describe tokens, context windows, and their practical consequences.
• Construct clear, specific prompts that produce useful output.
• Apply role, context, task, format, and constraint elements in prompt construction.
• Apply few-shot prompting by supplying worked examples.
• Apply structured reasoning prompts for multi-step problems.
• Iterate and refine prompts systematically based on output quality.
• Evaluate AI output for accuracy, relevance, completeness, and bias.
• Verify factual claims in AI output against authoritative sources.
• Apply AI tools to text tasks: drafting, summarizing, rewriting, and translating.
• Apply AI tools to code tasks: generation, explanation, debugging, and documentation.
• Apply AI tools to data and analysis tasks appropriately.
• Describe retrieval-augmented approaches and why grounding output in sources matters.
• Use an AI system's application programming interface at an introductory level.
• Describe prompt injection and other adversarial risks at an introductory level.
• Apply data privacy and confidentiality requirements when using AI services.
• Describe bias in training data and output, and its consequences.
• Describe intellectual property and attribution questions raised by generative output.
• Apply academic and workplace policy governing AI use.
• Document AI use transparently in submitted or delivered work.

Optional Outcomes

• Build a simple application integrating an AI model.
• Describe fine-tuning and when it is appropriate.
• Describe multimodal models and image generation.
• Describe evaluation methods for AI system output.
• Describe local and open-weight model deployment.
• Describe the labour and environmental costs of AI systems.

Major Topics

Required Topics

• What a language model is and how it generates
• Capabilities and inherent limitations
• Hallucination and confident error
• Tokens and context windows
• Prompt construction fundamentals
• Role, context, task, format, constraints
• Few-shot prompting
• Structured reasoning prompts
• Systematic iteration and refinement
• Evaluating output quality
• Verification against authoritative sources
• Text applications
• Code applications
• Data and analysis applications
• Retrieval augmentation and grounding
• Working with an API
• Prompt injection and adversarial risk
• Data privacy and confidentiality
• Bias in data and output
• Intellectual property and attribution
• Academic and workplace policy
• Transparent documentation of AI use

Optional Topics

• Building an AI-integrated application
• Fine-tuning
• Multimodal and image generation
• Output evaluation methods
• Local and open-weight deployment
• Labour and environmental costs

Resources & Tools

• Provider documentation — the model providers publish prompting guides and API references, and they are free, current, and better than any textbook in a field moving this fast. Read the documentation for whichever system your course uses.
• Free tiers — most major systems offer free access sufficient for coursework. Check your institution's approved tools before using anything with course data.
• Python with the provider SDKs — the COP1000 prerequisite makes API work accessible, and API access is where the durable skill is.
• Ollama or LM Studio — free, and let you run open-weight models locally. Genuinely instructive: watching a smaller model fail shows you what the large ones are doing.
• Hugging Face — free model hosting, datasets, and documentation; the field's practical hub.
• NIST AI Risk Management Framework (nist.gov) — free, and the most useful sober treatment of AI risk available to a student.
• U.S. Copyright Office — free guidance on registration and AI-generated material; see the flag below.
• Your institution's academic integrity policy and, later, your employer's AI policy — the operative rules, and both are changing.
• Primary sources for verification — library databases, government data, and standards bodies. The verification habit is the course's most transferable output.

Career Pathways

• Software developer — AI-assisted development is now ordinary practice, and fluency is expected rather than remarkable.
• Data analyst — AI tools accelerate exploration; the judgement about what is trustworthy remains human.
• Technical writer and content developer — drafting and editing workflows have changed substantially.
• Business analyst and operations — process automation and document workflows.
• Customer support and service design — assistant systems and their supervision.
• Marketing and communications — content production at volume, with the accuracy problem attached.
• AI integration and solutions roles — building applications on top of model APIs; this is where the programming prerequisite pays.
• Quality, compliance, and AI governance — an emerging area as organizations adopt policy.
• Any knowledge role, indirectly. The realistic framing is that this is a general workplace competency rather than a job title.
• SOC codes 15-1252 Software Developers, 15-2051 Data Scientists, and 15-1299 Computer Occupations, All Other. See the honest note below on "prompt engineer" as a career.

Special Information

⚠⚠ These systems are confidently wrong — verification is the whole professional skill

The single most important content in the course, and the one that determines whether a student is safe to let near real work.

Language models generate plausible text; they do not check facts. The output is fluent, confident, well structured, and may be entirely fabricated — this is commonly called hallucination, and it is a property of how the systems work rather than a bug awaiting a fix.

• Fluency is not accuracy. The confident tone is generated the same way the content is, and it carries no information about correctness. This is precisely why the errors are dangerous: they do not look like errors.
• Citations and references are frequently fabricated. Plausible-looking sources — correct-sounding authors, real journals, invented articles — are a well-documented failure mode. Every citation must be checked to exist and to say what it is claimed to say. Lawyers have been sanctioned for filing briefs containing invented cases.
• Numbers, dates, and quantities are unreliable. Arithmetic and specific figures should be independently verified or computed by a tool.
• Errors cluster where you can least check them — outside your own expertise, which is exactly where the temptation to use the tool is strongest.
• Ask for sources and then verify them, rather than trusting output that includes them. Retrieval-augmented approaches that ground output in supplied documents reduce but do not eliminate the problem.
• Never submit or deliver output you have not verified and do not understand. Being unable to explain your own work in a meeting is worse than not having produced it.

The professional framing worth adopting: the model drafts; you are accountable. That division of responsibility does not change regardless of how good the tools become, and it is the reason this course belongs in a curriculum rather than being left to self-teaching.

⚠ What you paste in may leave your organization

The confidentiality issue that has produced real incidents and real policy.

• Text entered into a consumer AI service leaves your control. Depending on the service and the tier, it may be retained, reviewed, or used to improve the system. Pasting confidential material into one can be a disclosure.
• This matters more in some fields than others, and it matters legally in several. Patient information engages HIPAA; student records engage FERPA; client information engages professional confidentiality duties; and personal data engages Florida's data breach law, § 501.171, F.S., which this repository documents elsewhere.
• Employers have policies, and they vary sharply — some prohibit external AI services entirely, some provide sanctioned enterprise instances with contractual protections, some are silent. Ask before pasting anything, and treat silence as a reason to be conservative rather than as permission.
• Enterprise and API tiers frequently carry different data commitments than consumer interfaces. The distinction is real and worth understanding.
• De-identify where you can. Much of the useful work can be done on a redacted version of a document.
• Prompt injection is a genuine security concern once these systems process untrusted input — content in a document or a web page can contain instructions the model follows. Anyone building applications on these systems needs to understand it.

⚠ Bias, intellectual property, and the questions without settled answers

Content a responsible course covers even though — perhaps especially because — the answers are unsettled.

Bias: models learn from large text corpora and reproduce patterns in that data, including social biases. Documented effects include differential quality of output across dialects and languages, stereotyped associations, and uneven performance for different groups. The practical consequence is that output should be reviewed for it, particularly in anything affecting people — hiring material, evaluations, or public communication.

Intellectual property: this is genuinely unsettled and moving.

• The U.S. Copyright Office has taken the position that purely AI-generated material without sufficient human authorship is not copyrightable, while works with meaningful human contribution may be protected in that contribution. Registration guidance requires disclosure of AI-generated content.
• Whether training on copyrighted material is infringement is being actively litigated, with substantial cases pending. Nobody should represent the answer as settled.
• Output can resemble training data closely in some circumstances, which creates practical risk in commercial use.
• Terms of service vary on who owns output and what it may be used for; read them for the specific service.

Academic integrity: institutional policy is in flux and varies by instructor within a single institution. Using a prohibited tool is a violation regardless of how much it helped, and AI-detection tools are unreliable in both directions — producing false accusations and missing actual use — which is itself a reason to be transparent rather than to rely on not being caught. Ask, follow the stated policy, and document your use when permitted.

Rule 11 applies with unusual force to this entire section. Copyright law, institutional policy, employer rules, and the capabilities of the systems themselves are all changing. Verify current provisions — anything specific in this guide should be treated as a starting point rather than as current fact.

⚠ The honest career note: "prompt engineer" is a contested job title

Worth stating plainly so students calibrate expectations correctly.

• The standalone "prompt engineer" role was widely discussed and has not become a large, durable job category. The reasons are structural: models became better at interpreting ordinary instructions, prompting techniques diffused quickly, and the skill turned out to be an adjunct to domain expertise rather than a profession on its own.
• The durable version of this skill is different and more valuable — knowing what these systems are good and bad at, structuring a problem so a tool can help, verifying output, and integrating AI into a workflow responsibly. That is a general competency in nearly every knowledge role.
• Domain knowledge is what makes AI use effective. You cannot evaluate output in a field you do not understand, which means the tool amplifies expertise rather than replacing it.
• Programming skill compounds it. API access, building applications, and automating workflows are where the significant employment is, which is why the COP1000 prerequisite matters.
• Expect the specifics to date. Techniques taught this term may be unnecessary next year as models improve. Learn the reasoning, not the recipes.
• Be careful with claims on a résumé. "Familiar with AI tools" is credible; asserting expertise in a fast-moving field invites questions you should be able to answer.

How Florida course levels affect transfer

The first digit of an SCNS number denotes the year of offering, not transferability. Courses at the 1000 and 2000 levels transfer transparently between Florida public institutions, and 3000 to 4000 is unproblematic since both are upper division. The boundary that actually matters is 2000 to 3000, where lower-division credit generally cannot satisfy an upper-division requirement.

CAI1320C is 3 credits with an estimated 60 contact hours under the C-suffix combined lecture-and-laboratory convention. Expect applied project work — prompt development and iteration, output evaluation, verification exercises, and in stronger sections an application built against an API — rather than examination-heavy assessment.

Transfer requires the suffix check above, and one further caution specific to this prefix: CAI is new, and receiving institutions may not yet have an equivalent course or a settled place for it in a degree plan. Have any transfer evaluated in writing, and note that computing A.S. degrees are applied and do not carry the A.A.'s guaranteed junior-status transfer.