| name | landing-page-experiments |
| description | Generate A/B test backlogs and experiment ideas for a landing page. Outputs a scored, prioritized backlog of 10-20 tests grouped by element (hero, proof, form, CTA, pricing) and hypothesis type (clarity, trust, friction, urgency, specificity), each with a verbatim current state, specific variant copy, primary metric, guardrail, expected lift band, win-rate prior, min sample size estimate, effort estimate, and "why this might lose" failure mode. Always works against a fresh landing-page-summary. Reads prior CRO audit history so it never suggests fixes already shipped. Reads the brand kit if present so variant copy stays on-brand. Use when the user asks to "give me A/B tests", "what should we test", "test ideas for this page", "experiment backlog", "next sprint tests", or "after shipping the audit, what's next".
|
| argument-hint | <URL> [--focus hero|proof|form|cta|pricing|all] [--win-bias higher-clarity|higher-trust|lower-friction] [--traffic meta|google|tiktok|organic|email|retargeting] [--icp <description>] [--history-from YYYY-MM-DD] |
Landing Page Experiments
A/B test ideation. Output is a scored, prioritized backlog of tests for a specific landing page. Not an audit (that's optimize-landing-page). Not a recap. The user opens it and has a real sprint plan.
When to use this vs. optimize-landing-page:
optimize-landing-page answers "what's wrong with my page?" Once per page per quarter-ish. Returns top 3 fixes plus a short test queue appendix.landing-page-experiments answers "what should we test next?" Once per sprint, after a win, after a loss. Returns a deep backlog with element-by-element coverage, hypothesis-type grouping, and explicit "why this might lose" framing.
The two skills share the upstream landing-page-summary evidence layer but diverge in output shape and decision support.
Inputs
- URL (required).
--focus. Scope the backlog to one element or all (default). Options: hero, proof, form, cta, pricing, all.--win-bias. Adjusts scoring priors. higher-clarity weights clarity-of-message tests up. higher-trust weights social proof, guarantees, badges up. lower-friction weights form/CTA simplification up. Default: balanced.--traffic. Traffic source for variant suggestions (cold Meta vs retargeting vs search needs different tests).--icp. Target customer description.--history-from. ISO date. Look at prior CRO audits and prior experiment backlogs from this date forward to avoid repeating tests already suggested or fixes already shipped. Default: 90 days back.
If URL is missing, ask. If both --traffic and --focus are missing, proceed with defaults (all focus, no traffic-temperature adjustment) and note assumptions in the output.
Freshness Gate
Always invoke landing-page-summary fresh for the URL in this conversation. Never read a saved summary directly. Stale summaries silently corrupt test ideation just like they corrupt audits and brand kits.
If landing-page-summary is not installed, fall back to WebFetch with the exhaustive extraction prompt. Quality is degraded. Flag it in the backlog header.
Runtime Notes
- Use
TodoWrite to track phases when the full backlog is being built.
- Ask questions in chat. Do not wait for a tool prompt.
- Save the finished backlog to
./artifacts/experiments.<host>.<date>.md (the copy the user opens) AND to /agent/brain/experiments/<slug>--<domain>/<YYYY-MM-DD>.md (durable backlog history). Multiple backlogs per day get a -HHMM suffix.
- Hand the artifact back using a
file link widget.
Method
Phase 1 — Get the Page Spec (Fresh)
Invoke landing-page-summary at --depth standard. Read the result. The summary's H2 anchors are the data contract.
Phase 2 — Read Prior Context
For each of these, read what exists. Skip silently if absent.
- Prior CRO audits.
/agent/brain/cro-audits/<slug>--<domain>/*.md — read every audit since --history-from. Extract the "Top 3 Fixes" sections and note their status. Tests suggesting the same fix are pre-filtered out of the backlog (or marked "already shipped, skip").
- Prior experiment backlogs.
/agent/brain/experiments/<slug>--<domain>/*.md — read every backlog since --history-from. Extract test IDs and hypothesis summaries. New backlog should not repeat the same tests verbatim; if the same hypothesis applies, mark it as "carried forward from " and refresh the variant.
- Brand kit.
/agent/brain/brand-kit/_index.md — if a brand kit exists for the target brand, read the markdown source. All variant copy in the backlog must match the brand's voice rules, vocabulary, and anti-patterns. If no brand kit, note the gap and use Motion's HTML design system fallback for any visual variant specs.
Phase 3 — Generate the Backlog
Produce 10-20 tests. Each test is a complete decision record.
Test record schema:
test-NNN
Element: hero | above-fold-trust | body-section | form | pricing | final-cta
Hypothesis type: clarity | trust | friction | urgency | specificity | mechanism | outcome
Hypothesis: "{{Replacing X with Y will {{increase|decrease}} {{metric}} because {{causal claim}}.}}"
Current state: "{{verbatim quote from summary}}"
Variant: "{{specific replacement copy or structural change}}"
Primary metric: {{CVR / form completion / CTA click-through / scroll depth}}
Guardrail: {{don't-regress metric, e.g. bounce rate, time on page, downstream revenue}}
Expected lift: {{band: 1-3% | 3-7% | 7-15% | 15%+}}
Win-rate prior: {{percentage from win-rate priors table}}
Min sample size: {{N visitors per variant for MDE = lower bound of expected lift at 95% confidence, given baseline CVR if known or "assume 3% baseline"}}
Effort: {{<1h copy change | 1-4h dev | 1d | 1wk}}
Score: {{(median expected lift) x (win-rate prior) / (effort hours)}}
Why this might lose: {{the specific failure mode. Be honest.}}
Dependencies: {{tests this depends on or conflicts with, or "none"}}
Win-rate priors (industry-rough; adjust per --win-bias):
| Hypothesis type | Default win rate | Best-fit element |
|---|
| Clarity (WHO/WHAT/WHY in hero) | 30% | hero |
| Friction reduction (form fields, fewer steps) | 50% | form |
| Risk reversal (guarantees, free trial, no-CC) | 40% | hero, pricing, final-cta |
| Specificity (real numbers, named customers) | 35% | proof, hero |
| Trust signal placement | 25% | above-fold-trust, proof |
| Mechanism reveal (how it works) | 30% | body-section |
| Outcome focus (results over features) | 30% | hero, body-section |
| CTA copy (outcome-oriented over generic) | 20% | hero, final-cta |
| Pricing display (anchor, comparison, popular) | 30% | pricing |
| Urgency (only if truthful and material) | 15% | hero, final-cta |
Use --win-bias to multiply priors: higher-clarity x 1.3 on clarity, higher-trust x 1.3 on trust/risk-reversal/specificity, lower-friction x 1.3 on friction/CTA. Cap at 80%.
Coverage requirement:
- At least 2 tests per element when
--focus all (unless an element doesn't exist on the page).
- At least 1 test per hypothesis type when win-rate prior > 25%.
- At most 4 tests per element to avoid backlog bloat.
- Total 10-20 tests.
Filtering:
- Tests that duplicate prior audit Top 3 fixes are excluded unless the audit fix was rejected or the page changed since. Note the exclusion in the report's "Filtered" section.
- Tests that duplicate prior backlog tests are marked "carried forward" with the original test ID for traceability.
Phase 4 — Compose the Output
Get the current timestamp:
date --iso-8601=seconds
Write the backlog to both paths (durable + artifact). Use this exact shape:
# Landing Page Experiment Backlog - {{Brand}} {{Page}} ({{URL}})
_Generated: {{ISO ts}} - Source summary fetched: {{summary-ts}} - Focus: {{focus}} - Test count: {{N}} - Win bias: {{bias or "balanced"}}_
_Brand kit applied: {{yes (path) | no - using design-system fallback}}_
_Prior context: {{N audits, M prior backlogs read since {{--history-from}}}}_
## Top 3 to ship first
| Rank | Test | Element | Expected lift | Effort | Score |
| --- | --- | --- | --- | --- | --- |
| 1 | test-NNN | | | | |
| 2 | | | | | |
| 3 | | | | | |
## Backlog by element
### Hero
- test-NNN: {{one-line hypothesis}}
- ...
### Above-fold trust
- ...
### Body sections
- ...
### Forms
- ...
### Pricing
- ...
### Final CTA
- ...
## Backlog by hypothesis type
### Clarity
- test-NNN, test-NNN
### Friction reduction
- ...
### Risk reversal
- ...
### Specificity
- ...
### Trust signal placement
- ...
### Mechanism reveal
- ...
### Outcome focus
- ...
### CTA copy
- ...
### Pricing display
- ...
### Urgency
- ...
## Test details
### test-001 - [Element] [Hypothesis type] [Score]
- **Hypothesis:** ...
- **Current:** "{{verbatim from summary}}"
- **Variant:** "{{specific replacement}}"
- **Primary metric:** ...
- **Guardrail:** ...
- **Expected lift:** ...
- **Win-rate prior:** ...
- **Min sample size:** ... per variant (assumes baseline CVR {{X%}}, MDE {{Y%}}, 95% confidence)
- **Effort:** ...
- **Why this might lose:** ...
- **Dependencies:** ...
### test-002 - ...
[repeat for every test]
## Filtered out
Tests excluded because they duplicate a prior audit fix already shipped or a prior backlog test still queued. Listed for visibility.
- {{prior audit fix excluded: ...}}
- {{prior backlog test carried forward as test-NNN: ...}}
## Dependencies and sequencing
Tests that block or conflict with each other. If you ship test-A first, test-B's variant may need refresh.
- test-NNN -> test-NNN (sequencing reason)
- test-NNN vs test-NNN (conflict, only run one)
Phase 5 — Index Hygiene And Drift Tracking
-
Update /agent/brain/experiments/_index.md with one row per page, most recent backlog first.
Row format:
- {{Brand}} {{page name}} - `<slug>--<domain>` - latest backlog {{date}}
- URL: {{URL}}
- Latest backlog: /agent/brain/experiments/<slug>--<domain>/<YYYY-MM-DD>.md
- Test count: N
- Top 3: {{test-001 summary}}, {{test-002 summary}}, {{test-003 summary}}
- Source summary fetched: {{summary-ts}}
- Brand kit applied: {{yes | no}}
-
Update /agent/INDEX.md with one entry per page (not one per backlog). Preserve created on update.
Use single-line |-separated format:
- path: /agent/brain/experiments/<slug>--<domain>/ | aliases: experiments, ab tests, test backlog, <brand> tests, <hostname> tests, landing page experiments | note: Experiment backlogs for <Brand> <page name>. One file per backlog at <YYYY-MM-DD>.md. Latest top test: <top-1 summary>. Source summary at /agent/brain/landing-pages/<slug>--<domain>.md. | created: <first-backlog date> | updated: <latest-backlog date>
Phase 6 — Surface Readout
After saving, return exactly 5 lines:
- Backlog: {{N tests}} for {{page name}} - top hypothesis: {{test-001 one-liner}}
- Quickest win: {{test with highest lift/effort ratio}}
- Biggest swing: {{test with highest expected lift}}
- Risk to watch: {{any test whose "why this might lose" surfaces a real downside}}
- Saved: {{artifact path}} - {{durable path}}
Hand off the artifact file via file link widget.
Rules
- Always re-fetch the page summary. Invoke
landing-page-summary fresh. Stale page state corrupts the backlog.
- Read prior audit history before generating. Never suggest a test for a fix that was already shipped. If unsure of status, mark the test "verify status: was this fix shipped?".
- Read the brand kit if it exists. All variant copy must obey the kit's voice, vocabulary, anti-patterns. No off-brand variant suggestions.
- Quote current state verbatim. Every test's
Current: field is exact copy from the page summary. No paraphrasing.
- Variant copy must be specific and brand-consistent. Not "improve the headline." Write the replacement headline.
- Win-rate priors are directional, not promises. State them. Do not claim "this will lift X%." Use the band.
- Min sample size estimate must show the baseline assumption. "Assumes 3% baseline CVR, MDE 0.5%, 95% confidence, two-tailed."
- Effort estimates honest. Distinguish copy-only changes (<1h) from dev work (1d+) from design-required (1wk+).
- Score formula consistent.
(median expected lift band as decimal) x (win-rate prior as decimal) / (effort hours as float). Higher is better.
- "Why this might lose" is required on every test. Surfaces the failure mode honestly. If you can't write one, you don't understand the test well enough yet.
- No dark patterns. No fake scarcity, no manipulative loss-aversion that isn't truthful, no fabricated testimonials, no deceptive CTAs.
- No "test everything" backlogs. 10-20 tests max. A 50-item backlog is a 0-item backlog.
- Filter, don't duplicate. Tests already shipped via prior audit fixes get excluded with a note. Tests still queued from prior backlogs get carried forward by ID.
- Cite the summary section (e.g. "from summary § 3 Hero") so the user can trace evidence.
- Save durable + hand off. Artifact copy is what the user opens this turn. Durable copy under
/agent/brain/experiments/ is what future Runneth can recall.
- Update both indexes. Local
_index.md and global /agent/INDEX.md on every backlog.
No Routine
This skill does not install a routine. Experiment ideation runs when the user wants it. The LP weekly refresh routine (from landing-page-summary) flags experiment backlogs whose pages changed materially, prompting the user to regenerate when relevant.
