Marketing · Frontier · updated 2026-10-23
Ad copy and creative variant agent
An agent reads a CSV of existing ads with performance, finds weak ones, and generates replacements that respect character limits (30 for headlines, 90 for descriptions) using two sub-agents. A Figma plugin then swaps copy into design templates to produce up to 100 creative variants at once.
Building block of the Self-improving ad creative loop.
1Is this my problem?
Anthropic's growth marketing was a non-technical team of one. This let them test far more ad variations without more people.
You'll recognise this if…
- Few ad variants each month.
- Character limits break uploads.
- Copy writing takes hours.
2Is this solution worth it?
Reported results
- Anthropic: ad copy work went from about 30 minutes per ad to 30 seconds; the Figma plugin generates up to 100 variations per batch (source)
Quoted from the linked source; results vary.
Your ROI, with the assumptions shown
Prefilled with illustrative assumptions (6 h/week at $60/h, 70% automated, build $5,000). Change them to your numbers.
Limits and when not to build it
Skip it if: Tiny ad budgets.
- Character limits and brand/legal rules must be checked in code, not trusted to the model.
- Keep a log of hypotheses and results or you will re-test the same ideas.
3How does it work?
One small internal tool does this for you. The same job, before and after:
Before · by hand
- Ads performance CSV export: someone notices
- Find underperforming ads
- Headline sub-agent (≤30 chars) (by hand)
- Description sub-agent (≤90 chars) (by hand)
- Figma plugin builds variants
- Manual review → upload CSV
Ad copy and creative variant agentIllustrative
- Starts automatically: ads performance csv export
- Find underperforming ads
- AI: Headline sub-agent (≤30 chars)
- AI: Description sub-agent (≤90 chars)
- Figma plugin builds variants
- You: Manual review → upload CSV
✦ AI step · ● human checkpoint · ✓ automatic
Generating many on-brand variants that obey hard constraints, informed by what performed before (a simple memory of past experiments).
- Headline sub-agent (≤30 chars)
- Description sub-agent (≤90 chars)
Your tool connects to
Under the hood: flow, architecture and vendors
- When: Ads performance CSV export
- Find underperforming ads
- Headline sub-agent (≤30 chars)
- Description sub-agent (≤90 chars)
- Figma plugin builds variants
- Manual review → upload CSV
Architecture
Vendors commonly used
Maturity ladder
What each level adds. Most teams start at level 1 or 2 and climb. Highlighted: where this build sits.
- Level 1Basic automation
Trigger: ads performance csv export. Fixed rules then find underperforming ads, figma plugin builds variants, manual review → upload csv. No AI involved.
- Level 2AI-assisted
An LLM handles the judgement step (headline sub-agent (≤30 chars)) and a person reviews every output before it is used.
- Level 3Autonomous agent
The agent runs end to end on its own and only stops for exceptions. Every ad is reviewed before upload.
- Level 4Self-learning system
Becomes part of the self-improving ad creative loop: A memory log of hypotheses and results is loaded into context next round, so the system learns which angles work instead of starting fresh.
Level 4 lives in the Self-improving ad creative loop: 6 stages with a learning loop.
Seen in the wild
- Anthropic: how Anthropic uses Claude in marketing
- Anthropic: how Anthropic teams use Claude Code (PDF)
Links checked 2026-10-23. We describe these builds from public material; we are not affiliated with these companies.
4Who can build it?
Matched on services, platforms and industries agencies list on their own sites, and on their published projects.
GoodspeedLondon, United KingdomWhy matched: Marketing automation.
Related project: n8n Automation for a Tax Workflow Platform — How we built production-ready n8n automations connecting a tax-workflow platform to HubSpot, Jira and the tools US accountancy firms already…
Published price: From $10,000
MakeitfutureCluj-Napoca, RomaniaWhy matched: Marketing automation.
Related project: How to Automate Email Campaigns with N8N — Make.com to n8n migration case study: Email campaign automation with memory-optimized workflows.
No published price
Connex DigitalUniversal City, United StatesWhy matched: Marketing automation.
Example project: Automating Casting Agency Communications: A Case Study — ChallengeA Los Angeles casting agency struggled to manage communications across multiple TV show productions.
No published price
Nex AutomationsAhmedabad, IndiaWhy matched: Marketing automation.
Example project: Five-Stage AI Article Pipeline on Make.com, Airtable and OpenAI — How an editorial team that rejected one-shot AI writing now runs a five-stage pipeline with a human gate at every stage. Make.com, Airtable,…
Published price: From $5,000
SynkrAIAhmedabad, IndiaWhy matched: Marketing automation.
No published price
Why matched: Marketing automation.
No published price
Projects, tools and prices come from each agency's own website (crawled October 2026). Not paid placements.
Part of the starter pack: Marketing team of one.
5What should I do next?
Questions
What does a "ad copy and creative variant agent" automation do?
An agent reads a CSV of existing ads with performance, finds weak ones, and generates replacements that respect character limits (30 for headlines, 90 for descriptions) using two sub-agents. A Figma plugin then swaps copy into design templates to produce up to 100 creative variants at once.
Which tools can I use to build it?
Common choices are Claude Code, Google Ads, Figma, Meta Ads, Google Sheets. Teams often use Claude Code, Google Ads, Figma together; the right mix depends on what you already pay for. n8n, Make and Zapier can all handle the workflow steps.
How much does it cost to build?
As a rough estimate, agencies often quote $2k–$8k for a first version of this workflow, plus monthly tool and AI usage fees. Your quote depends on how many systems are involved and how clean your data is.
How much time does it save?
Our estimate for a small team is about 4–10 hours a week. This is a rough guide, not a measured result; measure your own baseline before you start.
Where does AI actually help?
Generating many on-brand variants that obey hard constraints, informed by what performed before (a simple memory of past experiments).
Who actually runs this?
Public examples include: Anthropic: how Anthropic uses Claude in marketing; Anthropic: how Anthropic teams use Claude Code (PDF). Reported: Anthropic: ad copy work went from about 30 minutes per ad to 30 seconds; the Figma plugin generates up to 100 variations per batch.
Where should a person stay in the loop?
Every ad is reviewed before upload.
What usually goes wrong?
Character limits and brand/legal rules must be checked in code, not trusted to the model. Keep a log of hypotheses and results or you will re-test the same ideas.
Related workflows
Scheduled scrapes of competitor pricing pages, sites, ads and news are diffed against the last run; an LLM explains what changed and why it might matter, and a weekly digest goes to Slack..
Community questions into content and alertsMonitor relevant subreddits for questions and sentiment about your category; cluster recurring questions and turn them into answer posts or FAQ entries, and alert the team to threads worth a genuine reply..
Coding agent as a marketing team (skills library)Load an open-source library of marketing "skills" (copywriting, SEO audits, CRO, pricing pages, email) into a coding agent such as Claude Code, Codex or Cursor, and run them against your own site and data..
Programmatic SEO from your integration or data catalogGenerate one useful page per integration, use case or app pair from structured data you already have (supported triggers, actions, templates), with AI only filling descriptive gaps..