Headline Variant Tester

When to use this prompt

When you have a piece of content and a working headline that’s “fine but not great.” Or when you want to A/B test a headline but can’t write past two variants without them all sounding the same. Headlines do a disproportionate amount of the engagement work — getting them right is one of the highest-leverage edit passes in content production.

Use this for blog post titles, email subject lines, ad copy headlines, LinkedIn post hooks, podcast episode titles, conference talk titles. The framework adapts.

The prompt

<role>Editor specializing in headline variation and ranking.</role>

<task>Generate 5 distinct headline variants for the content below and rank them on three dimensions (specificity, curiosity, buyer-fit). Recommend one for the primary headline and one for an A/B test against it.</task>

<inputs>
<content_summary>
[2-4 sentences summarizing what the content actually says. Include the central claim, who it's for, and the takeaway.]
</content_summary>
<format>[blog post / email subject / ad headline / LinkedIn hook / podcast episode / conference talk]</format>
<buyer>[WHO IS READING THIS, e.g., "B2B marketing directors at mid-market SaaS"]</buyer>
<existing_headline>[YOUR CURRENT WORKING HEADLINE, IF YOU HAVE ONE]</existing_headline>
<character_limit>[OPTIONAL: max characters; defaults: blog 60, email 55, LinkedIn 100, ad 30]</character_limit>
</inputs>

<instructions>
1. Generate 5 headline variants. Each variant must be a distinct angle, not five rephrasings of the same angle. The five angles to cover (one variant each):
   - **Direct claim**: states the most provocative claim from the content as the headline
   - **Specific data**: leads with a specific number, percentage, or named entity
   - **Curiosity gap**: opens a question or unresolved tension the content answers
   - **Outcome promise**: names what the reader gets from reading
   - **Counter-take**: pushes against a common belief in the buyer's category
2. For each variant, score 1 to 5 on three dimensions:
   - **Specificity**: how concrete the headline is. "Improve your SEO" scores 1; "How a 30-prompt audit found 47% of citations were going to a 2-year-old comparison page" scores 5.
   - **Curiosity**: how strongly the headline pulls the reader to read more. Generic info scores 1; an unresolved tension scores 5.
   - **Buyer fit**: how clearly the headline signals to the named buyer that this is for them. Off-target audience scores 1; immediate signal scores 5.
3. For each variant, write one sentence on its strongest quality and one sentence on its weakest.
4. Recommend a primary headline and an A/B test variant. The two must be from different angles (do not pick "Direct claim" and "Specific data" if they're saying the same thing).
5. Constraints:
   - Stay under <character_limit>.
   - Do not invent specifics not implied in <content_summary>. Mark any number or named entity that isn't in the summary as [VERIFY].
   - Do not use clickbait patterns ("You won't believe...", "This one trick..."). They erode trust.
   - Avoid overused frame words like "Ultimate", "Complete", "Best", "Top 10" unless the content genuinely is one.
</instructions>

<output_format>
| # | Angle | Headline | Chars | Spec | Curi | BFit | Total |
|---|-------|----------|-------|------|------|------|-------|
| 1 | Direct claim | ... | XX | N | N | N | N/15 |
| 2 | Specific data | ... | XX | N | N | N | N/15 |
| 3 | Curiosity gap | ... | XX | N | N | N | N/15 |
| 4 | Outcome promise | ... | XX | N | N | N | N/15 |
| 5 | Counter-take | ... | XX | N | N | N | N/15 |

**Per-variant notes:**
1. Strongest: ... | Weakest: ...
2. Strongest: ... | Weakest: ...
3. ...
4. ...
5. ...

**Recommended primary:** Variant N — [reason]
**Recommended A/B variant:** Variant M — [reason this is the right opposing angle]

**[VERIFY] flags:** [Any number or named entity in any headline that isn't supported by the content_summary]
</output_format>

How it works

The five-angle structure is the highest-leverage move. Most headline brainstorming produces five rephrasings of the same angle (“Why X matters,” “What X means,” “Understanding X”) which all score similarly because they’re testing the same idea. Forcing distinct angles produces variants that actually test different hypotheses about what the buyer wants.

The three-dimension scoring (specificity, curiosity, buyer-fit) is intentional. Specificity alone produces dense but unenthusiastic headlines. Curiosity alone produces hook-y but vague headlines. Buyer fit alone produces relevant but boring headlines. Scoring all three together lets the model recommend a variant that does all three jobs.

The “primary + A/B” recommendation forces the model to commit to a pair, not a list. Lists are easy to dismiss; commitments are useful. The constraint to pick from different angles ensures the A/B test is actually testing different hypotheses about the buyer, not testing two near-identical variants against each other.

The forbidden-pattern list (clickbait, “Ultimate”, “Complete”) catches lazy frames. Frontier models default to these patterns when not told otherwise; the explicit blacklist keeps the headlines honest.

Example output

#	Angle	Headline	Chars	Spec	Curi	BFit	Total
1	Direct claim	Stop measuring traffic. Measure who’s recommending you.	56	4	4	5	13/15
2	Specific data	47% of B2B AI citations now skip the URL entirely	49	5	4	4	13/15
3	Curiosity gap	Why your dashboard says “fine” while pipeline breaks	52	3	5	4	12/15
4	Outcome promise	The 50-prompt audit that surfaces hidden citation share	54	4	3	4	11/15
5	Counter-take	AI Overviews don’t kill traffic. They kill the metric.	53	3	4	5	12/15

Recommended primary: Variant 1 — strongest combined score and the angle most novel to the buyer. Recommended A/B variant: Variant 3 — opposite angle (curiosity gap vs direct claim) tests whether the buyer responds better to a stated solution or an opened question.

[VERIFY] flags: Variant 2’s “47%” — confirm the number is supported by your data before publishing.

Variations

• Subject line A/B mode: For email subject lines specifically, generate 7 variants instead of 5 and add an “open rate prediction” dimension based on length, specificity, and curiosity.
• Conference talk mode: Constrain to 35 characters, add a “memorability” dimension (will this be remembered after 50 other talk titles), and recommend a single primary with no A/B (talks don’t A/B test).
• Refresh mode: For an existing headline that’s underperforming, add the current performance data (CTR, open rate) and ask the model to identify which dimension the current headline is weak on, then optimize specifically for that dimension.