Case Study 01
Turning vague ideas into useful AI outputs
A practical prompt engineering study exploring how context, constraints, and output structure affect text and image generation in GenAI Studio.
- Role
- Product designer & developer
- Focus
- Text + image prompting
- Method
- Controlled prompt comparison
01
The challenge
Generative AI can produce fluent content from a short instruction, but fluent does not always mean useful. Early tests in GenAI Studio produced broad copy and visually acceptable images that lacked audience, purpose, and a consistent direction.
Research question
How much does a structured prompt improve control over an AI result compared with a simple request?
02
Approach
I tested one variable: prompt quality. For each experiment, the task and output type stayed the same while the prompt changed from a short request to a structured brief. I reviewed the outputs using five practical criteria rather than treating personal preference as the only measure.
Relevance
Does the result answer the actual brief?
Specificity
Does it include the requested details?
Structure
Is the result organised and easy to use?
Tone & style
Does it match the intended audience and mood?
Consistency
Do repeated generations stay close to the brief?
03
Experiment one — text
Goal
Create a LinkedIn post about completing a Generative AI course.
Initial prompt
Write a post about learning AI.
Structured prompt
Act as a career coach. Write a 150–200 word LinkedIn post for recruiters and IT graduates about completing a Generative AI course. Use a confident but humble tone. Start with a short hook, share three lessons as bullet points, and end with a thoughtful question.
| Criterion | Initial prompt | Structured prompt |
|---|---|---|
| Audience | Not defined | Written for recruiters and IT graduates |
| Structure | General paragraphs | Hook, three lessons, closing question |
| Tone | Generic and promotional | Confident, reflective, and personal |
| Length | Unpredictable | Focused within the requested range |
Finding: defining the audience and format made the largest difference. The role added perspective, while the word range prevented unnecessary detail.
04
Experiment two — image
Goal
Create a portfolio image of a developer working in Johannesburg.
Initial prompt
A woman working on a laptop.
Structured prompt
Young woman developer working on a laptop in a bright Johannesburg co-working space, eye-level 50mm photograph, subject placed using the rule of thirds, warm late-afternoon window light, shallow depth of field, realistic detail, calm and optimistic mood.
| Criterion | Initial prompt | Structured prompt |
|---|---|---|
| Subject | A generic laptop user | A young woman developer at work |
| Setting | Unspecified office | Bright Johannesburg co-working space |
| Composition | Model-decided framing | Eye-level, rule-of-thirds framing |
| Visual direction | Inconsistent | Warm light, calm mood, realistic finish |
Finding: concrete visual language worked better than subjective adjectives. Subject, setting, camera angle, composition, lighting, and mood gave the model a coherent shot list.
05
What did not work
More detail was not always better.
Prompts with conflicting styles or too many instructions produced muddled results. Exact word counts were also treated as guidance rather than guaranteed rules.
- Over-prompting: competing instructions reduced clarity.
- Ambiguous language: words like “beautiful” gave little control.
- False certainty: polished text still required fact-checking.
- Variation: the same prompt could still produce different results.
06
Lessons learned
- 01
Treat prompts as specifications. A useful prompt explains the goal, audience, context, constraints, and expected output.
- 02
Change one thing at a time. Controlled iteration makes it easier to understand why an output improved or declined.
- 03
Use domain-specific language. Format terms help text models; composition and lighting terms help image models.
- 04
Keep human review in the loop. Prompt engineering improves direction, but it does not guarantee accuracy or quality.
07
Outcome & next steps
The experiments showed that prompt engineering is less about finding “magic words” and more about communicating intent clearly. The structured prompts produced outputs that were easier to evaluate, edit, and reuse in a real workflow.
The next version would run repeated generations for each prompt, score results against a fixed rubric, track prompt versions, and compare model performance using a larger test set.