Agentic Video Ads with Claude Code: Inside the AdSights Ads Framework

1. Monday — strategist writes a half-page brief (hooks, offer, product shots, brand colors)
2. Tuesday — dev partner or agency scaffolds the project and binds the brief to composition props (or a Claude Code skill does this from the brief doc)
3. Wednesday night — batch render runs: four hook × offer cells × two sizes = eight MP4s in a shared folder
4. Thursday — media buyer uploads to Ads Manager with consistent naming; strategist sets placement filters for Reels-only reads
5. Friday — pull thumbstop and hold from the Meta metrics walkthrough; kill cells below ~10% hold at 15s÷3s before increasing budget

Nothing here requires the strategist to learn Remotion APIs. They still own hooks, offers, and kill/scale rules — the framework removes the "one export per sprint" bottleneck.

A concrete workflow (before vs after)

Before (typical performance team):

1. Strategist writes a brief with three hook angles
2. Designer builds one timeline in After Effects (~2–4 days)
3. One MP4 exported; hook B and C wait for the next sprint
4. Media buyer uploads manually; no structured metadata

After (Ads Framework + optional Claude Code skills):

1. Strategist writes the same brief
2. Skill reads the brief and proposes six variants (two hooks × three offer lines) bound to composition props
3. CLI batch-renders all six at Reels + Feed sizes overnight
4. QA skill checks duration, safe zones, and burned-in captions
5. Media buyer receives a structured output folder with naming conventions intact

The strategist's job does not disappear — hooks, offers, and test design still require human judgment. What changes is how many cells you can run before fatigue sets in.

Example variant matrix (one brief → six renders)

Each cell is one composition render — not a timeline re-export. After upload, score hooks with the benchmark guide segmented by placement; kill variants below ~10% hold at 15s÷3s on cold traffic before scaling spend.

What the shipped concepts cover

The framework ships 8 ad concepts × 12 format sizes (Feed, Reels, Stories, YouTube pre-roll, and more). Concepts include:

• Product showcase — hero product, feature callouts, CTA
• Before/after — transformation narrative for DTC and fitness
• Testimonial quote — customer voice with on-screen attribution
• Sale promo — deadline-driven offer framing
• Talking-head UGC style — creator-style delivery without reshooting every variant
• Stat reveal — number-forward proof points
• Comparison vs — side-by-side competitive framing
• Listicle — multi-point value props with staggered reveals

Brand tokens (colors, typography, spacing) codegen from your config so re-skinning a variant matrix does not mean rebuilding frames.

What agents actually automate

Based on the skills shipped in the framework workspace:

The goal is not "AI makes ads." It is repeatable variant throughput so your creative testing framework gets enough cells to reach significance.

Getting started (non-technical path)

1. Browse the storefront — Frameworks hub lists Ads Framework, Email Framework, and Claude Skills bundles with pricing and licensing
2. Read setup docs — /frameworks/ads-framework/setup walks installation and first render
3. Run one vertical slice — scaffold with create-adsights-ads, render a single concept at one size, confirm the MP4 in Ads Manager
4. Add Claude Code skills (optional) — Ads Skills for Claude documents the brief → variant → render loop
5. Pair with measurement — use Hook Rate Benchmark and ThruPlay Rate Benchmark so variant volume produces learnings, not just files

If you already have an agency or dev partner on Node/Remotion, they can own layers 1–2 while your team owns briefs and test design.

Where measurement still matters

Agentic generation increases variant volume — which makes measurement discipline more important, not less. Pair the framework with:

• Hook Rate Benchmark — score whether new hooks clear thumbstop bands
• ThruPlay Rate Benchmark — score engagement depth on Meta
• How to run a Meta creative test — experiment design so variant volume becomes evidence
• Video Creative Metrics Benchmarks 2026 — cited ranges for hook, hold, and ThruPlay

High render throughput without experiment design is just faster noise.

Who this is for

Strong fit:

• Performance teams already running structured creative tests
• In-house teams with a technical partner (or agency) comfortable with Node projects
• Brands scaling localized or personalized video variants across many sizes

Weak fit:

• One-off brand films with no iteration loop
• Teams expecting a Canva-style template picker with zero project setup
• Organizations that cannot version-control creative assets in git

Common questions from media buyers

"Do I still upload to Ads Manager myself?" Yes — the framework outputs files and metadata; upload workflows stay the same unless you wire your own automation.

"Can I change the hook without reopening After Effects?" Yes — hooks are composition props. Update the text binding, re-render, done.

"What if we don't use Claude?" Use the CLI and engine directly. Skills are an accelerator, not a requirement.

"How does this connect to our benchmark stack?" Render more variants → test with pre-defined sample sizes → score winners with the analyzers above. The framework solves production throughput; your testing playbook solves inference.