How Orcool Builds Performance Creative at Scale with Anthropic's Claude

A look inside the signal-to-concept pipeline behind a weekly creative engine for one of the world's largest consumer mobility platforms.

The bottleneck nobody talks about in performance marketing

Ask any performance marketing leader where their growth is capped, and the honest answer is rarely the media budget. It's the creative.

Modern paid acquisition runs on volume and velocity. A single market might burn through dozens of ad variations a week before fatigue sets in. Multiply that across countries, languages, and cultural contexts, and the math breaks. Teams either slow down and ship fewer, safer creatives, or they flood the channel with variations that miss the local nuance that actually drives a tap. Neither path scales.

This is the problem Orcool was built to solve. And it's where we work alongside one of the most demanding creative operations in the world: one of the world's largest consumer mobility platforms, with a mobile-first product and a performance marketing engine that never stops.

The challenge: creative velocity without losing the local thread

Our partner's situation is one most growth teams will recognize, just at a larger scale. They needed:

• Volume. Enough fresh, distinct creative concepts every week to keep paid channels from fatiguing across dozens of active markets.
• Cross-market testing. The ability to test ideas in parallel across very different regions, rather than betting everything on a single hero concept built for one audience.
• Cultural nuance. Creative that reads as native in each market, not English-first ideas mechanically translated into other languages. In consumer mobility, the difference between a concept that converts and one that falls flat is often a culturally specific moment of trust, safety, or convenience.

Producing this by hand is possible. Producing it every week, at this breadth, without burning out a creative team, is not. The team needed a system that could generate at machine scale while reasoning at human depth.

Orcool's approach: a signal-to-concept pipeline

Most AI creative tools today take a URL and spit out an ad. That's not what we do. Orcool runs a signal intelligence pipeline that moves from raw market evidence to scored, production-ready creative concepts.

It works in three stages. First, signal intelligence: we ingest competitor advertising, platform trends, user reviews, and audience research, and decompose them into structured signals about what is actually working in a given market right now. Second, concept generation: those signals are recombined into distinct creative concepts, each scored against the brand's own performance lens before a human ever sees it. Third, production: approved concepts move into our creative production pipeline and come out as finished, channel-ready video.

The intelligence layer that makes the first two stages work is built on Anthropic's Claude.

Claude's role, component by component

We use Claude, built by Anthropic, as the reasoning core across the most judgment-heavy parts of the pipeline. Five components stand out.

Competitor ad decomposition. Claude reads competitor creative and breaks it down into its working parts: the hook, the emotional promise, the format, the cultural cue, the call to action. This turns a wall of competitor ads into a structured map of what's resonating in each market, which becomes raw material for new ideas rather than something to copy.

Concept generation with scoring. Claude generates distinct creative concepts from those signals and evaluates each one against a brand-specific scoring framework, so the team starts from a ranked shortlist instead of a blank page. In a typical weekly cycle this scores thousands of concepts per brand and surfaces the top few dozen to ship, each tied back to the signals that inspired it.

Script writing. For concepts that move forward, Claude drafts the actual scripts, including the dialogue and on-screen beats, in a voice tuned to the target market rather than a generic translation.

Review and insight mining. Claude reads through large volumes of user reviews and market feedback to surface the recurring tensions, delights, and objections that should shape the next round of creative. This is long-context work: the value comes from holding thousands of data points in view at once and finding the pattern.

Agentic orchestration via a custom MCP server. This is the piece we're proudest of. We built a custom server on the Model Context Protocol (MCP) that exposes our entire Orcool Studio environment to Claude as a set of tools. Claude doesn't just produce text we paste somewhere; it operates the platform: pulling signals, creating concepts, scoring them, and moving work through the pipeline. MCP turns Claude from an assistant into an operator inside our own system.

The results

Because this work runs inside an active, performance-driven account, the specific media metrics belong to our partner. What we can share is the operational shift.

The pipeline scores thousands of creative concepts a week per brand against that brand's lens, and ships the top few dozen across 10+ markets and languages, each one traceable back to the market signal that inspired it. The team moved from a place where creative was the bottleneck to one where there's a deep, ranked backlog of ideas to test. And across markets, we've seen measurable improvements in creative hit rate — more of what ships actually performs, because the concepts are grounded in real signal and scored before production rather than after.

The deeper result is structural. Creative velocity stopped being a function of how many hours the team could work, and started being a function of how good the signals going in were.

Why Claude, specifically

We've tested the alternatives. Three things made Claude the right reasoning core for this system.

Non-English market nuance. Claude writes and reasons in target-market voice rather than translating English ideas. For a product that lives in dozens of countries, this is not a nice-to-have.

Long-context reasoning. The review-mining and signal-decomposition steps depend on holding enormous amounts of context at once and reasoning over all of it. Claude's ability to sustain quality across very long inputs is what makes the insight layer trustworthy.

Agentic reliability. When a model is operating your platform through MCP rather than just chatting, reliability is everything. Claude follows multi-step instructions, uses tools predictably, and recovers gracefully, which is what lets us trust it with real production work rather than supervised drafts.

What's next

We're extending the same architecture: deeper feedback loops where live performance data flows back into the scoring lens, more markets onto the pipeline, and a tighter loop between signal and finished creative. The thesis stays the same. Performance creative shouldn't be a bottleneck. With the right signals and Claude as the reasoning core, it becomes a renewable resource.

Orcool is an AI-native creative studio building the signal-to-concept-to-production pipeline for performance marketing teams. We build on Anthropic's Claude.