Why Your AI Prompts Aren't Working
You've tried ChatGPT. You've tried Midjourney. But your results feel generic, off-brand, or just... okay. The issue isn't the tools—it's that most prompts are written like casual questions rather than precise specifications.
When you ask an AI "write me a job description," you get a job description. When you ask "write a job description for a mid-level UX researcher at a 50-person startup, emphasizing research methodology over management, with a tone that feels collaborative rather than corporate," you get something actually usable.
The difference is specificity. And it's learnable.
The Four Layers of a Strong Prompt
Think of prompt engineering like writing a detailed creative brief. Each layer adds clarity and reduces the AI's guessing.
Layer 1: The Core Task Start with what you actually need. Not "write about travel" but "write a 200-word budget travel guide for backpackers visiting Southeast Asia on $30 per day, focusing on lesser-known cities."
Layer 2: Context and Constraints Tell the AI what matters. Budget constraints, audience knowledge level, format requirements, length, tone, and any hard rules. Examples:
- "Assume the reader has no design experience"
- "Must fit in a Twitter/X post"
- "Avoid mentioning competing products"
- "Written for a 14-year-old audience"
Layer 3: Style and Voice Direction AI tools are surprisingly good at mimicking writing style when you're specific. Instead of "make it professional," try: "Written in the direct, no-filler style of Paul Graham's essays" or "Conversational and sarcastic, like a friend explaining this over coffee."
Layer 4: Output Format Specify exactly how you want the result structured. Bullet points, numbered steps, a table, a narrative, JSON, code blocks—precision here saves you editing time later.
Real Examples That Work
Image Prompts (Midjourney, DALL-E)
Weak: "A coffee shop"
Strong: "Overhead shot of a Scandinavian-style coffee shop interior, natural wood tables, single espresso cup in lower left frame, afternoon sunlight streaming through large windows, shot on a Leica M6 film camera with Portra 400 film, warm color grading, 4:3 aspect ratio"
The strong version includes: subject, camera angle, style reference, framing detail, specific location in frame, photographic equipment reference, color tone, and aspect ratio.
Text Prompts (ChatGPT, Gemini)
Weak: "Write me a resume for a marketing job"
Strong: "Write a resume for a marketing coordinator position at a B2B SaaS company. The candidate has 3 years of social media experience, event coordination background, and some HubSpot knowledge. Keep it to one page, use a modern format with short bullet points, emphasize metrics (engagement rates, event attendance numbers), and make the skills section highlight CMS and analytics tools. Tone should be confident but not overselling."
Notice: role specificity, candidate background, format constraints, emphasis areas, and tone direction.
Prompt Engineering Strategies That Actually Work
- Use role-play framing: "You are an experienced UX researcher. How would you approach testing this feature?" often produces better results than asking directly.
- Provide examples in your prompt: "Here are three subject lines that worked well [examples]. Write five more in this style." Examples anchor the AI's output.
- Chain requests instead of piling them on: First ask for raw ideas, then ask to refine, then ask to shorten. Multi-step prompts often outperform mega-prompts.
- Specify what to avoid explicitly: "Don't use corporate jargon" or "Avoid mentioning price" gives negative constraints that help.
- Use temperature and creativity settings: Most tools let you adjust how "creative" vs. "consistent" the output is. For factual content, lower is better. For brainstorming, higher works.
Testing and Iteration
Your first prompt won't be perfect. Strong prompt engineers treat it like A/B testing:
- Write a detailed first prompt
- Generate output
- Identify what's wrong (too formal? too long? missing a detail?)
- Adjust one or two elements
- Test again
Don't change everything at once—you won't know what actually worked.
Common Mistakes to Avoid
Asking for too much at once. If you want a 10-section strategic plan, a tagline, and three variations of each, break that into separate prompts. Overloading a single prompt dilutes focus.
Being vague about "quality." "High quality" means nothing. Instead: "Well-researched with at least three sources, factual accuracy verified, written for a PhD audience."
Not accounting for the tool's strengths. ChatGPT is excellent for writing and reasoning. Midjourney excels at complex visual styles. Gemini handles real-time information better. Match your task to the tool's actual abilities.
The Iterative Refinement Loop
The best prompts often come from running something rough, seeing the output, and thinking: Oh, I should have mentioned...
That insight becomes part of your next version. Over time, you develop intuition for what AI tools "understand" and how to talk to them.
Final Thought
Prompt engineering isn't magic—it's communication. You're learning to be clear in a way machines can act on. The tools themselves are capable of remarkable work. What changes everything is treating your prompt like a specification rather than a question. Explore ready-to-use AI prompts on Nohaya PromptAi to see how experienced prompt engineers structure requests across different tools and use cases.