ab-copy-writer ยท diff
git:20260504.0d54b6b to git:20260808.7dfb5c7
159 added, 88 removed. Audit A to A.
---
name: ab-copy-writer
- description: Generate A/B copy variants for any piece of marketing copy with a hypothesis per variant. Use when the user says "A/B test this", "write variants", "give me options", "test different angles", "copy variants", "ab copy", "test this headline", "multiple versions", or wants to test copy before publishing. Produces 3-5 variants per copy element with angle, hypothesis, and expected audience response. Reads brand voice and ICP from knowledge/.
- reads:
- - knowledge/brand/voice.md
- - knowledge/icp/personas.md
- - knowledge/markets/positioning.md
- - knowledge/learnings.md
- writes:
- - output/ab-copy/
+ description: Generate A/B copy variants grounded in Eugene Schwartz's awareness and market-sophistication stages and John Caples' tested-headline discipline. Every variant declares its angle, the awareness stage it targets, and a falsifiable hypothesis, and the skill recommends which pair to test first and what delta counts as signal. Use when the user says "A/B test this", "write variants", "give me options", "test different angles", "copy variants", "ab copy", "test this headline", "multiple versions", "which version is better", or wants to test copy before publishing. For reviewing copy that already exists, see copy-review. For sample size and stopping rules, see growth-experiment. For diagnosing a page rather than its words, see page-cro. For full paid ad sets by platform, see ad-campaign-writer.
+ metadata:
+ grounded_in:
+ - "Tested Advertising Methods - Caples"
+ - "Breakthrough Advertising - Schwartz"
+ reads:
+ - knowledge/brand/voice.md
+ - knowledge/icp/personas.md
+ - knowledge/markets/positioning.md
+ - knowledge/learnings.md
+ writes:
+ - output/ab-copy/
---
# ab-copy-writer
- Produces testable copy variants. Every variant has a distinct angle, a hypothesis for why it will outperform, and a predicted audience response. Stops marketers from shipping one version and guessing.
+ Produces testable copy variants and argues for a specific test. Most variant generators produce five versions of the same idea and leave the choice to taste. This one selects angles from the reader's awareness stage, prioritises the element with the most leverage, and states what result would prove it wrong.
- ## When to use
+ ## The two frameworks
- - "Write 3 variants of this LinkedIn post"
- - "Give me A/B options for this email subject line"
- - "Test different angles for our landing page hero"
- - "I need to A/B test this ad headline"
- - "Write copy variants for our webinar title"
- - "Give me 5 versions of this CTA"
+ ### 1. Awareness and sophistication (Eugene Schwartz, *Breakthrough Advertising*, 1966)
- ## Inputs needed
+ Schwartz's argument is that copy does not have a quality score in the abstract. It has a fit with where the reader already is. Two axes:
- 1. **The copy element**: what needs variants (headline, subject line, hero, CTA, LinkedIn hook, ad copy, webinar title, email opener)
- 2. **The source**: existing copy to riff on, or a brief description of what it should accomplish
- 3. **Number of variants**: default is 3, max is 5
- 4. **Channel**: where this copy will run (LinkedIn, email, landing page, ad, webinar registration page)
- 5. **What's being tested**: angle, tone, format, CTA, or length (ask if not clear)
+ **Awareness: how much the reader already knows.**
- ## Angles this skill tests
+ | Stage | The reader | Copy must |
+ |---|---|---|
+ | Unaware | Does not know they have the problem | Name the problem before selling anything |
+ | Problem-aware | Feels the pain, does not know solutions exist | Articulate the pain better than they can |
+ | Solution-aware | Knows solutions exist, not that yours does | Show the mechanism and why it is faster or surer |
+ | Product-aware | Knows your product, not convinced | Prove it, with specifics and social proof |
+ | Most aware | Ready, needs a reason to act now | State the offer and the deadline |
- For each variant, pick a distinct angle from this list. Never use the same angle twice in one set:
+ The common failure is writing product-aware copy for problem-aware traffic. It reads as a good ad that nobody responds to.
- | Angle | Description | When it wins |
+ **Sophistication: how tired the market is of this claim.**
+
+ | Stage | Market state | What still works |
|---|---|---|
- | **Outcome-led** | Leads with the specific result | When audience is ROI-focused and numbers-driven |
- | **Pain-led** | Opens on the problem, not the solution | When audience is in active pain and aware of it |
- | **Speed-led** | Emphasizes time-to-value | When audience is impatient or under pressure |
- | **Contrarian** | Challenges a common belief | When audience is saturated with standard messaging |
- | **Social proof** | Leads with a customer result or name | When audience trusts peers more than vendors |
- | **Curiosity gap** | Creates a knowledge gap that demands closing | When audience is information-hungry |
- | **Specificity** | Extremely specific number or detail as the hook | When vague claims are common and specifics stand out |
- | **Fear of missing out** | What happens if you don't act | When cost of inaction is high and under-communicated |
- | **How-to** | Practical, instructional framing | When audience wants actionable takeaways |
- | **Founder/operator voice** | Written as a practitioner to a practitioner | When audience distrusts brand voice |
+ | 1 | First to market | State the claim plainly. Nothing else needed |
+ | 2 | Competitors making the same claim | Make the claim bigger or more specific |
+ | 3 | Claims no longer believed | Lead with the mechanism, how it works |
+ | 4 | Mechanisms competing | Make your mechanism easier, faster, or surer |
+ | 5 | Everything sounds the same | Stop claiming. Identify with the reader's experience |
+ Sophistication is why a contrarian angle is mandatory in a crowded category and wasteful in a new one. Ask for it, or infer it from `knowledge/markets/competitors.md`.
+
+ ### 2. Tested headline discipline (John Caples, *Tested Advertising Methods*, 1932)
+
+ Caples ran split-run tests at BBDO to measure actual response rather than opinion, and the book is the origin of the discipline this skill supports. Three findings that change how we test:
+
+ 1. **The headline carries most of the leverage.** Changing the headline moves response far more than changing body copy. So test the headline first, and never burn a test cycle on body copy while the headline is untested.
+ 2. **Headline types are not equal.** Caples grouped them as news, self-interest or benefit, and curiosity, and found curiosity weakest **on its own**. A curiosity angle that carries no benefit is the most common way a clever variant loses.
+ 3. **Specific beats general.** A concrete number or detail consistently outperforms the same claim stated broadly.
+
+ ## Element priority (Caples)
+
+ Test in this order. Do not test element 3 while element 1 is unresolved.
+
+ | Priority | Element | Why |
+ |---|---|---|
+ | 1 | Headline, subject line, LinkedIn hook, ad headline | Highest leverage by a wide margin |
+ | 2 | Offer and CTA wording | Changes what is being asked, not just how |
+ | 3 | Opening line, sub-headline | Decides whether the headline's promise is kept |
+ | 4 | Body copy, length, formatting | Real but small. Test last |
+
+ If the user asks for body-copy variants while the headline has never been tested, say so and offer to test the headline first.
+
+ ## The angle table
+
+ Each angle is mapped to the awareness stage where it can work, the sophistication level it suits, and its Caples type. Never use the same angle twice in one set, and never pick an angle whose stage does not match the traffic.
+
+ | Angle | Works at awareness | Suits sophistication | Caples type |
+ |---|---|---|---|
+ | **Pain-led** | Problem-aware | 1-2 | Self-interest |
+ | **Outcome-led** | Solution / product-aware | 1-2 | Self-interest |
+ | **Specificity** | Solution / product-aware | 2-3 | Self-interest + news |
+ | **Mechanism-led** | Solution-aware | 3-4 | News |
+ | **Speed-led** | Solution-aware | 2-4 | Self-interest |
+ | **Social proof** | Product-aware | 3-4 | News |
+ | **Contrarian** | Unaware / problem-aware | 4-5 | Curiosity + self-interest |
+ | **Curiosity gap** | Unaware / problem-aware | 4-5 | Curiosity (never alone) |
+ | **How-to** | Problem / solution-aware | 2-3 | Self-interest |
+ | **Identification** | Any, strongest at 5 | 5 | Self-interest |
+ | **FOMO / deadline** | Most aware | any | News |
+ | **Founder voice** | Any | 4-5 | Identification |
+
+ **Hard rule from Caples:** a curiosity or contrarian variant must also carry a benefit. Curiosity that pays off in nothing is the most reliable loser in the set, and it is usually the variant that reads cleverest in the document.
+
## Process
- 1. **Load context.** Read `knowledge/brand/voice.md` (for tone and style rules) and `knowledge/icp/personas.md` (for what this audience responds to). If either is missing, ask the user for the relevant context directly.
+ 1. **Load context.** Read `knowledge/brand/voice.md` for tone and banned words, `knowledge/icp/personas.md` for the language this audience uses, and `knowledge/markets/positioning.md` for the claim the copy must carry. If voice is missing, continue and mark the output generic rather than stopping.
- 2. **Check `knowledge/learnings.md`** for any past A/B test results or copy performance data. If a previous test showed "outcome-led beats pain-led for our ICP", factor that in and note it.
+ 2. **Check `knowledge/learnings.md`** for past test results. If a prior test settled an angle for this ICP, say so and weight accordingly rather than re-running a decided question.
- 3. **Understand the goal.** Ask if not provided: "What action should this copy drive? And what's the one thing the audience should feel or believe after reading it?"
+ 3. **Establish the two stages before writing anything.** Ask, or infer and state the inference:
+ - **Awareness:** where is this traffic coming from, and what do they already know? Cold LinkedIn feed is problem-aware at best. A pricing page is product-aware. A retargeting ad is most aware.
+ - **Sophistication:** how many competitors make this same claim already? Check `knowledge/markets/competitors.md`.
- 4. **Generate variants.** For each:
+ Both go in the output header. A variant set without a declared stage is a set of guesses, and every hypothesis below it is unfalsifiable.
- ```
- ## Variant [A/B/C/D/E]: [Angle name]
+ 4. **Confirm the element and its priority.** If the requested element is priority 3 or 4 and priority 1 is untested, flag it once.
- **Copy:**
- [the actual copy - complete and ready to use]
+ 5. **Select angles.** Filter the table to the declared awareness stage, then prefer angles matching the sophistication level. Default 3 variants, maximum 5. If the user supplied existing copy, include one variant that keeps their angle and sharpens the execution, so there is a baseline.
- **Angle**: [name from the angles table]
- **Hypothesis**: [1-2 sentences on why this angle should outperform the others for this specific audience]
- **Predicted response**: [how the ICP is likely to feel reading this]
- **Best for**: [which segment of the ICP or which stage of the funnel this variant suits most]
- **Watch for**: [what signal would tell you this variant is winning - clicks, replies, saves, conversions]
- ```
+ 6. **Write each variant:**
- 5. **Recommend the test setup**:
+ ```
+ ## Variant [A-E]: [angle]
- ```
- ## Recommended test setup
+ **Copy:**
+ [complete, ready to paste]
- **Test**: [Variant A] vs [Variant B] (start with the two most different angles)
- **Why these two**: [1 sentence on why this pairing gives the clearest signal]
- **Sample size needed**: [rough guidance - "at least 500 impressions per variant before reading results"]
- **How long to run**: [1-2 weeks for LinkedIn organic, 7 days minimum for paid, 3+ sends for email]
- **Primary metric**: [clicks / replies / saves / open rate / conversion - pick one]
- **How to read results**: [what delta counts as a win: ">10% difference in CTR is signal, <5% is noise"]
- ```
+ **Angle:** [from the table]
+ **Awareness stage:** [stage this is written for]
+ **Caples type:** [news | self-interest | curiosity + benefit]
+ **Hypothesis:** [why this should outperform for THIS ICP, referencing the stage]
+ **Would be proven wrong if:** [the observation that would kill it]
+ **Watch:** [the one metric that reads it]
+ ```
- 6. **Self-check each variant**:
- - Distinct angle from every other variant - no two variants are the same approach with different words
- - Matches `knowledge/brand/voice.md` style rules (no em dashes, no banned phrases, correct POV)
- - Uses language from `knowledge/icp/personas.md` where relevant
- - Hypothesis is specific to the ICP, not generic ("this will outperform because our ICP is measured on activation rate, not awareness, so outcome-led copy speaks directly to their KPI")
+ 7. **Recommend the test.** Pair by stage, not by contrast:
- 7. **Save** to `output/ab-copy/<DD-MM-YYYY>-<copy-element-slug>.md` with frontmatter:
- ```yaml
- ---
- format: ab-copy
- element: [headline|subject-line|hero|cta|hook|ad]
- channel: [linkedin|email|landing-page|ad|webinar]
- variants: [number]
- angles: [list of angles used]
- created: DD-MM-YYYY
- ---
- ```
+ ```
+ ## Recommended test
- 8. **Offer follow-ups**:
- - "Run `/retro` after the test and log which variant won and why. This gets appended to `knowledge/learnings.md` so future copy benefits from the result."
- - "Run `/content-writer` or `/linkedin-post` to build out the winning variant into a full piece."
- - "Want me to write the full ad set or email sequence using Variant [X] as the control?"
+ **Test:** Variant [X] vs Variant [Y]
+ **Why this pair:** [the one assumption they isolate]
+ **Element:** [name] (priority [N] of 4)
+ **Declared stages:** awareness [stage], sophistication [level]
+ **Primary metric:** [one]
+ **Signal threshold:** [what delta is worth acting on for this volume]
+ **Before launching:** run `/growth-experiment` to set sample size and the stopping rule.
+ ```
+ 8. **Save** to `output/ab-copy/<DD-MM-YYYY>-<element-slug>.md` with frontmatter carrying `element`, `channel`, `awareness-stage`, `sophistication-level`, `angles`, `variants`, `created`.
+
+ 9. **Offer follow-ups:** run `/retro` after the test so the winning angle lands in `knowledge/learnings.md`, or `/content-writer` to build the winner out.
+
+ ## Pairing logic
+
+ Testing the two most different variants is a weak default, because a loss tells you nothing about which of the several differences caused it.
+
+ 1. **Isolate one variable.** Two variants at the same awareness stage with different angles test the angle. Two variants of the same angle at different stages test the stage. Pick which question is being asked and hold everything else steady.
+ 2. **When the stage itself is uncertain**, that is the more valuable test. Run problem-aware against product-aware copy first. The answer redirects every future piece for this channel.
+ 3. **Never test three or more simultaneously** on low volume. Two variants need enough traffic already.
+
## Copy length by channel
| Channel | Element | Length |
|---|---|---|
| LinkedIn | Hook (first 2 lines) | 15-25 words |
| LinkedIn | Full post | 80-220 words |
| Email | Subject line | 6-10 words |
| Email | Preview text | 8-12 words |
- | Email | Opener (first sentence) | 10-20 words |
| Landing page | H1 | 5-10 words |
| Landing page | Sub-headline | 15-25 words |
- | Google ad | Headline | 30 chars max |
- | Google ad | Description | 90 chars max |
- | LinkedIn ad | Intro text | 150 chars max |
+ | Google ad | Headline | 30 characters |
+ | Google ad | Description | 90 characters |
+ | LinkedIn ad | Intro text | 150 characters |
| Webinar | Title | 8-12 words |
| CTA button | Text | 2-5 words |
+ Verify current platform limits before shipping paid copy. These change.
+
+ ## Self-check
+
+ - Awareness and sophistication are declared in the header, not implied
+ - Every angle is valid for the declared awareness stage
+ - No two variants share an angle
+ - Every curiosity or contrarian variant also carries a benefit
+ - Every hypothesis names the ICP and the stage, not "this will perform well"
+ - Every variant has a "would be proven wrong if" line
+ - The recommended pair isolates exactly one variable, and the output says which
+ - The element's priority is stated, and a priority 3-4 request with an untested headline was flagged
+ - Variants differ in approach, not in word choice
+ - Voice rules from `knowledge/brand/voice.md` are respected, or their absence is declared
+
## Rules
- - Never produce variants that are only cosmetically different (swapping one word is not a different angle).
- - Never use an angle that `knowledge/learnings.md` has already shown underperforms for this ICP without flagging: "Past data shows [angle] underperformed. Including it here for completeness, but Variant X is the higher-confidence bet."
- - Every hypothesis must reference the ICP specifically. "This will perform well" is not a hypothesis. "Our ICP (Head of Growth, measured on activation rate) will respond to specificity because they are data-fluent and distrust vague ROI claims" is.
- - Always recommend starting the test with the two most different variants, not adjacent ones. Adjacent variants give weak signal.
- - If the user provides existing copy, always include one variant that is a direct evolution of their version (keep their angle, sharpen the execution) alongside genuinely different angles. They may want a baseline.
+ - **Never produce cosmetic variants.** Swapping a word is not an angle.
+ - **Never invent a statistic** to make a specificity variant concrete. Tag `[NEEDS INPUT: real figure]`. A fabricated number in a winning variant ships to the whole audience.
+ - **Never predict a percentage lift.** Give direction and confidence, and let the test produce the number.
+ - If `knowledge/learnings.md` shows an angle already lost for this ICP, do not silently include it. Flag it or leave it out.
+ - If the copy's problem is the offer rather than the words, say so and point at `/positioning-doc`. Testing headlines on a weak offer optimises the wrong thing.
+
+ ## Related skills
+
+ - `/ad-campaign-writer` applies the same Schwartz stages to full paid ad sets by platform
+ - `/growth-experiment` sets sample size and stopping rules before the test runs
+ - `/copy-review` improves a single piece rather than generating alternatives
+ - `/page-cro` when the page underperforms for reasons beyond its words
+ - `/retro` records which angle won, so `knowledge/learnings.md` improves the next set
+ - `/brand-context` creates the voice, persona and competitor files this reads
+
+ ## What this skill cannot know
+
+ - Whether the traffic actually arriving at this copy matches the awareness stage you declared
+ - Current platform character limits, which change without notice
+ - Whether a variant that won last quarter still wins, since audiences fatigue
+ - The sample size this test needs. Run `/growth-experiment` before launching