Architecting Agentic Customer Avatars: Implementing Multi-Agent Focus Group Workflows in Claude Code

In the current landscape of Large Language Models (LLMs), a significant hurdle for developers and marketers is "sycophancy"—the tendency of models to act as "yes men," reflexively validating user prompts regardless of their actual merit. When using Claude or similar LLMs for market validation, this inherent bias can lead to catastrophic failures in product-market fit.

To solve this, I have implemented a custom skill within Claude Code (and compatible environments like Cursor and Codex) that utilizes an agentic orchestration pattern to simulate a high-fidelity, multi-agent focus group. This system does not rely on generic personas; instead, it constructs synthetic avatars derived from the feature extraction of real-world sales transcripts.

The Architecture of Synthetic Personas

The core innovation lies in moving away from zero-shot prompting toward a data-driven, composite avatar architecture. The system follows a structured pipeline: Ingestion $\rightarrow$ Feature Extraction $\rightarrow$ Clustering $\rightarrow$ Dossier Generation.

1. Data Ingestion and Feature Extraction

The workflow begins by pointing the skill at a repository of sales call transcripts (sourced via local directories or through the Model Context Protocol (MCP) from tools like Fireflies or Google Drive). The system initiates a "fan-out" pattern, deploying sub-agents to parse these transcripts.

These sub-agents are tasked with identifying and extracting specific high-signal features:

• Linguistic Markers: Exact vocabulary, phrasing, and semantic patterns used by customers.
• Objection Mapping: Specific friction points raised during the sales cycle.
• Pricing Sensitivity: Qualitative data regarding reactions to cost and value propositions.
• Decision-Making Hierarchy: Identifying stakeholders and influencers within the buying process.

The extracted data is stored in a centralized "quote bank," ensuring that the agents' responses are grounded in empirical evidence rather than hallucinated personas.

2. Composite Avatar Construction (Clustering)

To prevent overfitting to single individuals, the system employs a clustering logic. The skill groups customers based on shared buying behaviors and pain points. An avatar is only instantiated if there is sufficient data density—specifically, at least three distinct real-world customers must contribute to a single seat on the focus panel. This ensures that no agent is a 1:1 copy of an actual client, maintaining a level of abstraction that preserves privacy while maximizing representative accuracy.

Each resulting avatar is codified into a Markdown dossier. These dossiers contain the persona's identity, buying triggers, and historical objections. To ensure long-term consistency, each agent maintains a persistent position_file. This file is updated after every session, allowing for stateful memory across different focus group iterations.

The Multi-Agent Focus Group Protocol

The execution of a "focus group" follows a rigorous, multi-stage protocol designed to eliminate groupthink and maximize the discovery of divergent opinions.

Phase 1: Isolated Reaction & Gut Check

When the /focus_group command is invoked, the system spins up separate, isolated sub-agents for every seat on the panel. Crucially, these agents operate in complete isolation; Agent A has no visibility into the context or responses of Agent B. This prevents the "cascading agreement" common in standard chat interfaces.

The reaction occurs in two distinct layers:

1. The Gut Check: An immediate, low-latency assessment to determine if the input (landing page copy, email hook, etc.) captures attention.
2. The Spoken Take: If the gut check indicates a significant reaction, the agent generates a detailed qualitative analysis. If the initial reaction is dismissive, the agent provides only a brief "brush off," mimicking real-world human behavior.

Phase 2: The Moderator Agent and Re-engagement Loop

Once the isolated reactions are collected, a high-level Moderator Agent parses the aggregate output. The moderator's sole objective is to identify "sharpest disagreements"—points where the panel splits. It does not inject its own opinion; it acts strictly as an analytical layer for conflict detection.

In the second round of the protocol, only the "holdout" agents (those who disagreed with the majority or expressed significant friction) are re-engaged. These agents are presented with the opposing arguments. This allows us to test the robustness of a position: Would a person in this specific persona's position genuinely change their mind when presented with these counter-arguments?

Use Cases for Agentic Market Validation

This system provides three primary utility vectors for product development and marketing:

1. Messaging & Positioning Iteration: Testing landing pages, email hooks, and pricing structures against the panel before deployment. This allows for "tuning" a launch to attract Ideal Customer Profiles (ICPs) while intentionally repelling bad-fit prospects through specific linguistic triggers.
2. Concept Validation: Before committing engineering resources to a new feature, the panel can be queried to determine if the proposed solution addresses documented pain points in the transcript library.
3. Post-Mortem Failure Analysis: If a launch fails (e.g., high click-through rate but zero conversions), the exact assets can be run through the panel to identify which specific objection was triggered and why the value proposition failed to convert.

Implementation and Scalability

The skill is designed for modularity. By leveraging MCP, the system can scale its knowledge base dynamically. As new sales calls are added to the ecosystem, the skill automatically refreshes the dossiers. If a new cluster of data emerges that does not align with existing avatars, the system proposes the creation of a new seat on the panel, allowing the synthetic market to evolve in tandem with real-world market shifts.

For those looking to implement this within Claude Code, the installation involves adding the marketplace and the focus group skill via CLI commands, pointing the data source to your local or MCP-connected transcript repository.