The stack
The problem
Most LinkedIn ad programs stall for one reason that nobody names: they test variations of a single idea and call it testing. Five versions of 'our platform is faster' is not a test, it is the same bet placed five times with slightly different wording. When all five underperform, you have learned nothing except that this one idea did not land, and you have spent a month and a budget to learn it. The thing that actually moves performance is testing genuinely different strategic angles against each other: a cost-of-inaction angle versus a contrarian take versus a peer-proof angle versus a status-quo critique.
Generating a truly diverse angle slate by hand is slow and quietly sabotaged by the team's blind spots. People default to the angles they personally believe in, which is usually a feature-superiority angle because that is what they argue about internally all day. An AI primed with your positioning and, crucially, your customers' actual words will propose angles a single marketer would not reach for, and it will do it across several formats in minutes instead of over a planning week.
The catch, and it is a big one, is that generic AI ad copy is genuinely bad. Ask a model to 'write LinkedIn ads for our product' cold and you get interchangeable buzzword sludge. The entire difference between useless and useful output is grounding: feed the model your real positioning, your differentiators, your customers' verbatim language from reviews and call transcripts, and then constrain it to LinkedIn's actual format realities. Garbage in, confident garbage out.
The opinion that should drive the whole exercise: your highest-performing hooks are almost never phrases your marketing team wrote. They are phrases your customers used to describe their own pain, lifted verbatim. The model's job is not to be clever; it is to surface and structure the language that already resonates because real buyers already said it.
How it works
- Assemble an input pack: positioning, differentiators, the target persona, and real customer quotes
- Generate 8-10 genuinely distinct messaging angles, each with a different psychological premise
- Pick the 4-5 most distinct and expand each into hooks and full copy per LinkedIn format
- Spec a matching visual creative concept for each angle so copy and design reinforce each other
- Lay it out as a test matrix in Google Sheets, one row per variant with a results section
- Launch distinct angles first, kill losers, then expand hooks under the winning angle and feed winners back in
See it run
The playbook
Assemble the input pack before you prompt anything
Collect the raw material first, because the output quality is capped entirely by these inputs. Gather four things: your one-line positioning, your top three differentiators stated concretely, the specific buyer persona for this campaign (title, industry, what they actually care about and worry about), and a handful of direct customer quotes. The quotes are the secret ingredient. Pull them from G2 and Capterra reviews, from sales-call transcripts in Gong or Fathom, and from support tickets. Customer language gives the model phrasing that resonates instead of corporate speak.
If you are short on real quotes, use Perplexity to research how this persona describes the problem in their own words: search communities, forums, Reddit threads, and industry articles where practitioners complain about the status quo. You are mining for the verbatim phrases people use when no marketer is in the room.
Write all of this into one document you will paste as the input block. Spending twenty minutes here is the difference between a slate of sharp, testable angles and a pile of filler you throw away.
TipPull the exact phrases customers use to describe their pain, including the unflattering ones. 'I spend my whole Friday rebuilding the same forecast' is a better hook source than any feature list, and you cannot invent that phrasing from inside the building.
Generate 8-10 genuinely distinct angles
Open Claude and run the angle prompt with your input pack pasted in. The instruction that matters most is demanding genuinely different strategic premises, not reworded copy. Ask the model to name the psychological premise behind each angle so you can reason about your portfolio of bets the way you would reason about a portfolio of investments: are these actually different, or am I making the same bet four times?
Aim for eight to ten angles, then select the four or five most distinct to develop. The discipline is resisting the two angles that are secretly the same bet wearing different outfits. If 'we save you time' and 'we make your team more efficient' both made the list, they are one angle and you should cut one.
Read each angle's premise critically. A premise like 'quantify the hidden cost they are not counting' is a real, distinct strategic bet. A premise like 'emphasize our quality' is not a strategy, it is a vague hope, and angles built on it tend to underperform.
You are a B2B performance marketer planning LinkedIn ads.
POSITIONING: {{ONE_LINE_POSITIONING}}
DIFFERENTIATORS: {{TOP_3, stated concretely}}
TARGET PERSONA: {{TITLE, INDUSTRY, WHAT THEY CARE ABOUT, WHAT THEY WORRY ABOUT}}
REAL CUSTOMER QUOTES:
{{PASTE_3-5_VERBATIM_QUOTES}}
Give me 8 DISTINCT messaging angles. Each MUST use a different strategic premise. Draw from premises like: cost of inaction, contrarian take, status-quo critique, specific quantified outcome, identity/status, risk reduction, speed, peer/social proof, reframing the problem.
For each angle provide:
- Angle name
- Premise (one sentence on why it could work psychologically)
- Core claim it makes
- Which customer quote or pain it draws from
Rules:
- Do NOT give me variations of the same idea. Each angle must be arguable AGAINST the others.
- Use the persona's own language from the quotes where possible.
- Banned words: unlock, leverage, supercharge, seamless, game-changer, revolutionize, effortless.TipIf two angles feel similar, ask Claude directly: 'Are angles 3 and 6 actually the same bet? If so, replace one with a more distinct premise.' The model is good at catching its own redundancy when you make it adversarial.
Expand chosen angles into hooks and full copy
For each angle you kept, generate the actual LinkedIn ad copy across formats: the single-image intro text, a short-form variant, and a document/carousel headline set. Constrain the model to LinkedIn's real format constraints, the most important being that only roughly the first 140 to 150 characters of intro text show before the 'see more' truncation. The hook has to land its whole job above that fold, because most people never click 'see more.'
Generate two or three distinct hook variants per angle. This lets you test the angle and the hook as separate variables later, which is what disciplined testing requires. Hold the angle constant across its hook variants so that when you run the test, a difference in performance isolates the hook, not a confound of angle-plus-hook changing together.
Ban the patterns LinkedIn's audience punishes: the 'Are you struggling with...' opener, the engagement-bait question, the fake-vulnerability hook. B2B buyers on LinkedIn have strong pattern-recognition for ads and the tired formats get scrolled past instantly.
Take this angle: {{ANGLE_NAME + PREMISE + CLAIM}}
Persona: {{PERSONA}}. Product (one line): {{ONE_LINE}}.
Write LinkedIn ad copy:
1. THREE hook options. Each is the first line, must work and intrigue WITHIN the first 140 characters (before 'see more'). No 'Are you struggling with...', no engagement-bait, no clickbait.
2. Full single-image ad intro text (under 600 characters) built on the strongest hook. Specific and concrete throughout.
3. A 4-slide carousel outline, one line per slide, that tells this angle's story.
4. A clear CTA that matches the angle (a cost-of-inaction angle and a peer-proof angle should NOT use the same CTA).
Use the persona's own language. Banned words: unlock, leverage, supercharge, seamless, game-changer, revolutionize, effortless.TipPaste each hook into a plain text editor and count to 140 characters. If the intriguing part lands after that, the hook fails on mobile no matter how good the full line reads. Rewrite so the tension is fully present before the cut.
Spec a matching creative concept for each angle
Copy and visual have to reinforce the same angle, or you split your message against itself. Ask the model to describe the visual concept for each angle in enough detail that a designer or an image-generation tool can execute it: the visual metaphor, the on-image text (kept very short), the format, and the emotional tone. A contrarian angle needs a visually different look than a peer-proof angle; pairing a stark contrarian hook with a smiling-stock-photo visual neuters both.
This step bridges strategy to production. Instead of handing a designer a vague 'make us some ads,' you hand them an angle, the copy, and a concrete visual brief. That dramatically cuts revision cycles and keeps the creative on-strategy rather than drifting toward whatever stock imagery was easy to find.
Be explicit about what to avoid visually, because the default for B2B ads is generic stock photography that signals 'ad' and gets ignored. A stat card, a stark comparison, an annotated screenshot, or a simple bold-text treatment almost always outperforms a stock photo of people pointing at a laptop.
For this angle and copy: {{ANGLE + WINNING_HOOK}}
Describe the visual creative concept for a LinkedIn single-image ad (1200x627 or square 1080x1080):
- Visual concept / metaphor that reinforces THIS angle specifically (not a generic version)
- On-image headline text, under 8 words
- Format suggestion: stat card, side-by-side comparison, annotated screenshot, bold-text treatment, simple illustration
- Color and emotional tone (e.g. stark and serious for a cost-of-inaction angle)
- What to AVOID visually (be specific, e.g. 'no stock photos of people at laptops')
Keep it executable by a designer or an image generator. The concept MUST match the angle's premise.Build the test matrix in a sheet
Lay everything out in Google Sheets as a proper test matrix: one row per ad variant, with columns for angle, premise, hook variant, format, full copy, creative concept, target audience, and a results block (spend, impressions, CTR, CPL, conversions). This turns a pile of ideas into a disciplined test plan and, just as importantly, a permanent record you can learn from across campaigns instead of re-deriving the same lessons quarterly.
Launch the test the right way: start with a small set of distinct angles, one strong hook each, against the same audience and budget so the angle is the only variable. Let LinkedIn's algorithm gather enough data before judging (small B2B audiences accumulate clicks slowly, so do not kill a variant on two days of thin data). Then kill the clear losers and expand hook variants under the winning angle, which is the second, narrower test.
Close the loop. Once you have a winning ad and its real metrics, paste both back into Claude and ask it to generate more in that proven direction. You are compounding learning: each cohort makes the next one sharper because you are feeding real performance data back into generation, not starting from positioning every time.
TipAdd a 'learning' column you fill in when you kill a variant, in plain language: 'cost-of-inaction beat feature-superiority 2:1 on CTR with this persona.' In six months this column is worth more than any single winning ad, because it is your accumulated map of what this audience responds to.
What you get
A single test-ready angle with its premise, hook variants, intro copy, and matching creative brief, as one row of the matrix.
ANGLE: The hidden cost of manual planning
Premise: Quantifying wasted hours the buyer is not counting beats praising features, because it reframes inaction as the expensive choice.
Core claim: Ops teams quietly lose roughly 9 hours a week each to manual planning.
Drawn from quote: 'I spend my whole Friday rebuilding the same regional forecast.'
HOOK VARIANTS (each lands before 'see more'):
A: Your ops team isn't slow. Your planning process is.
B: 9 hours a week. That's what manual planning quietly costs each person.
C: The most expensive part of your ops process is the part nobody measures.
INTRO TEXT (built on Hook B):
9 hours a week. That's what manual planning quietly costs each person on your ops team, and it never shows up on a P&L. Most ops leaders assume their team is at capacity. Usually it's not the team. It's the hours lost rebuilding the same plans by hand, every week, across every region... [see more]
CTA: See where the 9 hours actually go ->
CREATIVE CONCEPT:
Format: stat card. A calendar-week grid with most cells greyed out and one block highlighted, big '9 hrs / week' overlaid.
On-image headline: 'Where the week actually goes.'
Tone: stark, serious, no people, no stock photography.
Avoid: smiling teams at laptops, abstract swooshes, anything that reads as a generic SaaS ad.
AUDIENCE: Ops & Supply Chain leaders, 200-2,000 employee logistics/manufacturing, US/UK.
RESULTS: [spend] [CTR] [CPL] [conv] [learning]Pitfalls to avoid
Variations masquerading as a testFive reworded versions of one claim is the same bet placed five times and teaches you almost nothing. Insist on genuinely distinct strategic premises per angle, and cut any two angles that are secretly the same bet.
No real customer languageWithout verbatim customer quotes as input, the model produces confident corporate filler. Ground every generation in how buyers actually describe their pain, mined from reviews, calls, and communities.
Ignoring the 'see more' truncationIf the hook's tension lands after the first ~140 characters, the angle never gets read on mobile, where most impressions happen. Write and count for the truncation point, not the full line.
Copy and creative mismatchA sharp contrarian hook paired with generic stock-photo creative neuters both halves of the ad. Spec the visual to reinforce the specific angle's premise, and avoid the default stock imagery that signals 'ad.'
Killing variants on thin dataSmall B2B audiences accumulate clicks slowly, so a variant judged after two days may just be under-sampled, not bad. Let the algorithm gather enough data before declaring winners and losers, or you will kill your eventual best performer early.