Short-form Video Ads
Playbook for 15-30 second video creative across platforms.
The first 2 seconds of video is a UX problem, not a creative one.
- Ship 6 variants per angle; Meta Advantage+ and PMax now reward volume over polish.
- Use Clay + Common Room to pull 3 signal layers before 11x writes a single script.
- Framer + Relume wins the landing page; Next.js on Vercel only when the stack needs custom logic.
- PostHog + Northbeam is the 2026 baseline; if you cannot read hold rate, you cannot scale.
- Kill 80% of briefs; keep 1 rule: 1 hook, 1 claim, 1 CTA, 15-30 seconds.
- 01Day 1: Build a 50-account target table and enrich with role and pain signals.
- 02Day 2: Draft 15 hooks and score them against a fixed rubric.
- 03Day 3: Cut 3 scripts and export 9:16 variants at 3 lengths each.
- 04Day 4: Launch 3 angle bundles on Meta Advantage+ and Google PMax against 1 landing page.
- 05Day 5: Wire attribution so creative_id follows the user end to end.
- 06Day 6: Query spend, hold rate, and revenue; kill any asset below target.
- 07Day 7: Feed winners back into the brief generator and produce the next 5 briefs.
1. Build the creative system before you build the edit
Short-form wins in 2026 when the system is engineered before the timeline. Clay should enrich 3 inputs per account: ICP fit, recent intent, and category pain; Common Room should add product or community signals; Attio keeps the source of truth. Feed those into 11x or Lindy to draft 10 hooks and 5 proof points per segment. Use Claude 4.5 Sonnet in n8n to score each hook against a prompt like: 'rate for specificity, believability, and pattern interruption, 1-5'. PostHog then ties each asset back to watched seconds, clicks, and downstream conversion.
- 01
Map 3 signal layers
Build a Clay table with firmographics, intent, and objections; enrich with Common Room and Apollo AI.
Prompt: 'Return 3 pain hypotheses for {{company}} using job posts, reviews, and page visits.' - 02
Generate angle families
Have 11x or Relevance AI create 5 angle families per ICP: fear, speed, proof, comparison, and teardown.
Prompt: 'Write 5 angles for finance buyers; each angle must include 1 claim, 1 metric, 1 objection.'
- 03
Score before production
Use Claude 4.5 Sonnet through n8n to rank hooks, then only brief the top 3 into Framer or Next.js pages.
SQL in Cursor: select angle, hook, cta from creative_scores where score >= 4.2 order by predicted_hold desc;
2. Script for retention, not for taste
A 20-second ad in 2026 needs 3 beats: hook in 2 seconds, proof by second 8, CTA by second 18. 11x can draft the first cut, but Lindy should personalise the opener by segment so the same edit feels native. Use Runway or Google Veo 3 for motion inserts, then cut in a 9:16 template inside Framer or a lightweight Next.js render flow. Cursor is for revops SQL only when you need to join spend, view-through, and CRM revenue. The rule: if the script cannot be read aloud in 12 seconds, it is too dense.
- 01
Write 3-line scripts
Keep each script to hook, proof, CTA; 11x drafts 10 versions, Claude 4.5 Sonnet trims to 3.
Prompt: 'Write a 15s script for {{ICP}} using one customer metric and one objection rebuttal.' - 02
Personalise the first frame
Use Lindy or Relevance AI to swap the first 3 words by segment and channel.
Prompt: 'For CFOs, replace the opener with a finance-specific pain and a number.'
- 03
Modularise the edit
Cut 5 reusable openers, 5 proof blocks, and 5 CTAs in Runway or Veo 3, then recombine.
Template: 'Hook A + Proof C + CTA 2' for {{campaign_name}}.
3. Test like a media buyer, not a designer
Meta Advantage+ and Google Performance Max want 6 to 12 variants per angle, not 1 polished cut. Launch 3 angles x 4 edits each, then let Northbeam read CPA, assisted conversion, and payback by creative ID. PostHog handles post-click behaviour; June tracks activation if the video sells a product trial. Use n8n to route winners into a fresh batch every 72 hours, and use Clay to create the next set of audiences based on who watched 50% plus. The winner is the angle with stable CPA after spend doubles, not the one with the prettiest cut.
- 01
Launch variant bundles
Ship 3 angles with 4 edits each across Meta Advantage+ and PMax.
Campaign rule: 1 angle = 1 audience = 4 cuts = 1 landing page.
- 02
Read incrementality
Use Northbeam plus PostHog to separate click-based lift from real revenue lift.
Prompt: 'Which creative IDs drove assisted revenue within 14 days?'
- 03
Refresh every 72 hours
Trigger n8n to pull winning signals and spin up the next 4 hooks via 11x.
Trigger: 'if CPA drops 15% and spend > $500, clone the asset family.'
4. Close the loop with measurement and RevOps
The 2026 playbook dies without clean attribution. Northbeam owns channel-level truth, PostHog owns product behaviour, and Attio or HubSpot AI owns lead and revenue objects. Use Claude Code or Cursor to write SQL joins between ad IDs, landing events, and pipeline outcomes; keep the query in Git so no one hand-edits a dashboard. MCP servers can expose creative metadata to the agent so 11x knows which proof point matched which segment. If a creative cannot be traced from impression to opportunity in under 5 clicks, it is not production-ready.
- 01
Unify IDs
Pass creative_id, campaign_id, and audience_id into PostHog and Attio at creation time.
Schema rule: creative_id must equal asset_name in Meta and Northbeam.
- 02
Join revenue
Use Cursor or Claude Code to build SQL joining spend, sessions, activation, and closed-won.
Query: select creative_id, sum(revenue) from attribution where date >= now() - interval '30 days';
- 03
Teach the agent
Expose the reporting table via an MCP server so 11x can propose the next brief from winners.
Prompt: 'Use the top 3 creative IDs and write the next 5 briefs for {{segment}}.'
What goes wrong
Making 1 premium edit and calling it a testing plan.
Ship 3 angles x 4 cuts; Meta Advantage+ needs volume, not taste.
Using generic hooks from old briefs.
Pull 3 live signals from Clay, Common Room, and Apollo AI before scripting.
Reading CTR without post-click data.
Join PostHog and Northbeam before deciding a winner.
Hand-editing reports in spreadsheets.
Move joins into Cursor or Claude Code and keep SQL in Git.
Waiting 2 weeks to refresh fatigue.
Trigger n8n every 72 hours and replace losers fast.
How many versions should I launch before judging a short-form concept?▾
Launch 12 variants minimum: 3 angles, 4 edits each. Meta Advantage+ and PMax need statistical shape, not a single polished file. Use Northbeam and PostHog to read the winner after spend doubles, not after 50 clicks. If one version is all you have, you do not have a test.
What is the best AI-agent setup for creative production?▾
Use Clay for enrichment, 11x or Lindy for first-draft scripts, and Claude 4.5 Sonnet in n8n for scoring and rewrite control. Add an MCP server so the agent can read creative metadata and performance tables. That stack beats manual briefs because it turns signals into scripts in one loop.
Should I build landing pages in Framer or Next.js?▾
Default to Framer + Relume for 80% of short-form campaigns. Use Next.js on Vercel only when you need custom logic, dynamic proof, or deeper experimentation. The rule is simple: if the page is mostly copy and sections, Framer wins; if it needs product-state rendering, use Next.js.
How do I know a creative is fatiguing?▾
Track CPA drift, 3-second hold rate, and frequency in Northbeam plus PostHog. If CPA rises 20% inside 10 days or hold rate drops 25% from the baseline, the asset is dying. n8n should auto-create replacements from the same angle family before the account starts leaking.
What should the CTA be in a 15-30 second ad?▾
Use 1 action only: start trial, book demo, or see pricing. If the CTA needs more than 1 clause, the script is overloaded. 11x should draft 3 CTA variants, and Claude 4.5 Sonnet should pick the one that matches the promised outcome and the landing page.
Get one system like this in your inbox each week
No fluff, no upsell. Unsubscribe anytime.