Bad prompt examples and how to fix them
Seven of the most common bad AI prompts, why each one fails, and a refined version that actually works.
1. “Write me a landing page”
Why it fails: No product, no audience, no offer, no proof, no constraint, no format. The model invents everything.
✅ Better prompt:
You are a senior conversion copywriter.
Product: {{product}} (one-sentence description)
Audience: {{audience}}
Main objection: {{main_objection}}
Promise: {{promise}}
Proof: {{proof}}
Task: Draft a landing page with hero, sub-hero, 3 benefit blocks, 1 objection-handling section, and a CTA.
Constraints: No buzzwords ("revolutionary", "unlock"). Max 50 words per block.
Decision rule: Pick the angle that most directly neutralizes the main objection. 2. “Give me content ideas”
Why it fails: No audience, no channel, no goal, no constraint. The result is a generic list anyone could write.
✅ Better prompt:
You are a content strategist.
Audience: {{audience}}
Channel: {{channel}} (e.g., LinkedIn, YouTube, newsletter)
Goal: {{goal}} (e.g., demand, education, retention)
Task: Give 10 content ideas that match the channel and goal.
Return: title, format, target keyword (if any), what the audience learns, why it converts.
Reject ideas that could appear on any brand's blog. 3. “Make this sound better”
Why it fails: "Better" is undefined. The model will pick a generic "more polished" voice instead of solving a real problem.
✅ Better prompt:
Task: Rewrite the text below for {{audience}}.
Goal: {{goal}} (e.g., increase clarity, reduce word count by 30%, shift tone to confident-not-pushy).
Constraints: Keep all numbers. Keep all proper nouns. No clichés.
Return: 3 alternative versions plus a one-sentence explanation of the strategic choice behind each. 4. “Write a cold email”
Why it fails: No sender, no recipient, no offer, no CTA, no proof. The result is a textbook template that gets ignored.
✅ Better prompt:
You are a B2B SDR.
Sender: {{sender_role}} at {{sender_company}} ({{sender_offer}})
Recipient: {{recipient_role}} at {{recipient_company}}
Trigger event: {{trigger}} (recent hire, funding, product launch, etc.)
Pain hypothesis: {{pain}}
Task: Draft a 70-word cold email.
Constraints: One pain sentence, one proof sentence, one CTA. No "I hope this finds you well".
Decision rule: Lead with the trigger event, not the offer. 5. “Create a marketing plan”
Why it fails: No business model, no budget, no horizon, no channel constraint. The model returns a template instead of a plan.
✅ Better prompt:
You are a senior marketing director.
Business: {{business_model}} ({{stage}})
Audience: {{audience}}
Budget: {{budget}} / quarter
Horizon: 90 days
Constraint: Only channels we can staff today: {{available_channels}}
Task: Build a 90-day marketing plan.
Return: north-star metric, 3 bets, 1 hypothesis per bet, weekly milestones, kill criteria.
Reject any channel we can't staff in week 1. 6. “Summarize this”
Why it fails: No audience, no length, no purpose. The result is a generic paragraph instead of a useful brief.
✅ Better prompt:
Task: Summarize the text below for {{audience}}.
Goal: {{purpose}} (decide / share / forward / archive).
Length: {{length}} (e.g., 80 words, 3 bullets, one paragraph).
Structure: lead with the decision the reader needs to make, then 3 supporting points, then 1 risk.
Reject filler sentences and restatement. 7. “Analyze my audience”
Why it fails: No source material, no segmentation, no decision the analysis should support. The model invents a persona.
✅ Better prompt:
You are a customer-research analyst.
Audience: {{audience}}
Source: {{source_material}} (interviews, support tickets, sales notes — paste inline)
Goal: {{goal}} (e.g., identify the strongest objection, find the trigger event, choose a hero promise)
Return: recurring objection (verbatim, with frequency), trigger event, verbatim language, top 3 buying-process bottlenecks, and the single recommendation for the next campaign. Want your own prompt graded?
Paste any rough prompt into the free bad prompt checker and Promptmkr will tell you what’s missing and turn it into a reusable prompt template.
Common questions
- What counts as a bad AI prompt?
- A prompt is bad when the model has to guess at something you already know. Missing audience, missing constraint, no output format, and no rule for ranking options are the four most common gaps. The output usually looks fine in isolation and turns out to be unusable the moment you try to apply it.
- Why does AI give generic answers?
- Because a generic answer is the safest response to an under-specified question. With no audience, no constraint and no ranking rule, the statistically safe output is the middle of the distribution — which reads as bland. Specificity in the prompt is what moves the output away from that middle.
- Is a longer prompt always better?
- No. Adding length without adding structure usually makes things worse, because more words give the model more competing signals to balance. The fix for a bad prompt is naming the job, the inputs and the output format, not padding it.
- How do I know if my prompt is the problem?
- Run it twice with no changes. If the two outputs differ in structure — not just wording — the prompt has not pinned down the format, and the model is making that decision for you each time. That is a prompt problem, not a model problem.