Every marketer now has access to the same AI models. The differentiator isn't the tool—it's how you use it.
Good prompting is a skill. Like writing or design, it can be learned. Here's what we've figured out.
The Core Principle: Context Is Everything
LLMs are pattern-completion engines. They predict the next token based on context. More context = better predictions.
Bad prompt:
Write ad copy for a skincare brand.
Good prompt:
Write 5 Facebook ad primary text options for a DTC skincare brand targeting women 25-40 who struggle with hormonal acne. Brand voice is clinical but warm—think dermatologist friend. Key differentiator is patented retinoid delivery system that reduces irritation. Goal is demo bookings. Keep each under 125 characters.
The second prompt constrains the output to something useful. Constraints aren't limitations—they're instructions.
The Framework: CONTEXT-TASK-FORMAT-CONSTRAINTS
Every prompt should answer four questions:
Context: What does the AI need to know to do this well?
- Who's the audience?
- What's the brand voice?
- What's been tried before?
- What's the competitive landscape?
Task: What exactly should the AI produce?
- One thing per prompt. "Write copy and analyze competitors" is two prompts.
- Be specific about the deliverable.
Format: How should the output be structured?
- Bullet points vs paragraphs
- Length limits
- Sections or headers
- Examples to match
Constraints: What should the AI avoid?
- Off-limits topics
- Tone to avoid
- Clichés to skip
- Formats that don't work
Specific Techniques That Work
Give Examples (Few-Shot Prompting)
Don't just describe what you want. Show it.
Here are two headlines that worked well for us:
- "Dermatologist-recommended. TikTok-approved."
- "Clear skin isn't genetic. It's strategic."
>
Write 10 more headlines in this style.
The model will match the pattern of your examples, not just your description of the pattern.
Role Assignment
Tell the model who it is before asking for output.
You are a senior direct response copywriter with 15 years of experience in DTC skincare. You specialize in clinical-but-approachable messaging.
This primes the model's "voice" in a way that generic prompts don't.
Chain of Thought
For complex analysis, ask the model to think step by step.
Analyze this landing page. First, identify the primary value proposition. Second, list all proof points. Third, evaluate the CTA strategy. Fourth, suggest three improvements with rationale.
Forcing intermediate steps improves output quality.
Iterative Refinement
Don't expect perfection in one shot. Plan for iteration.
Prompt 1: Generate 20 headline options
Prompt 2: These 5 are closest. Explain what makes each work, then create 10 variations of each.
Prompt 3: Now make them punchier. Under 6 words.
Three rounds of refinement beats one round of attempting perfection.
Common Use Cases
Ad Copy Generation
Context: DTC supplements brand targeting men 35-55 who want to maintain energy and focus as they age. Voice is confident and straightforward—no pseudoscience or miracle claims.
>
Task: Generate primary text options for Facebook ads promoting a nootropic stack.
>
Format: 10 options, each under 125 characters
>
Constraints: Avoid health claims that would violate ad policies. No "miracle" language. No references to competitors.
Competitive Analysis
I'll paste the homepage copy from three competitors. For each, identify:
1. Primary value proposition
2. Target audience (inferred)
3. Key differentiators claimed
4. Proof points used
5. Potential weaknesses or gaps
>
Then suggest positioning opportunities we could exploit.
Customer Research Synthesis
Here are 50 customer reviews from Amazon for a competitor's protein powder. Analyze for:
1. Common complaints (quote specific language)
2. Unexpected praise (what are they happy about that isn't the obvious stuff?)
3. Language patterns (how do they describe the product and problem?)
4. Opportunity spaces (unmet needs or desires)
Landing Page Feedback
Review this landing page copy. Evaluate against these criteria:
- Clarity of value proposition
- Specificity of proof points
- Strength of CTAs
- Objection handling
- Risk reduction (guarantees, social proof, etc.)
>
For each section, rate 1-10 and suggest specific improvements.
What AI Is Bad At
Prompting skill has limits. Some things don't work:
Strategic judgment: AI can analyze, but it can't decide. The decision still requires human judgment.
True novelty: AI recombines patterns. Genuinely new ideas are rare. Use it for variations, not inventions.
Brand intuition: The model doesn't feel your brand. It approximates based on descriptions. Final brand judgment stays with humans.
Current information: Training data has cutoffs. Real-time market conditions aren't in there.
The Workflow Integration
AI tools work best as accelerants in existing workflows, not replacements.
First draft acceleration: Human → AI → Human. Start with direction, get a draft, refine.
Variation generation: Create one version manually, use AI to generate variations.
Analysis scaffolding: AI does the busywork analysis, human does the synthesis and decision.
The tool is only as good as its operator. Learn to prompt well, and AI becomes a multiplier. Prompt poorly, and it's a distraction dressed up as productivity.