Episode 068 Part II Deep Dive: Design Systems as AI Context with TJ Pitre
Aired live: February 20, 2026

Introduction
In Part II of Episode 068, host Ben Callahan is joined again by co-host TJ Pitre—founder of Southleft, a front-end design development agency specializing in the intersection of AI and design systems—for a live community deep dive. This episode builds on Episode 068 Part I's exploration of the challenges that emerge when stochastic models try to keep the deterministic promises of a design system.

This week the question turned practical: what work needs to happen behind the scenes so your design system can serve as powerful, reliable AI context? Ben and TJ sent the question to 1,031 design system practitioners and received 184 responses. The community came ready to share—from MCP servers as structured sources of truth, to agentic feedback loops that validate component output against documentation, to honest debate about where Storybook fits in an AI-native workflow.

Show Notes
00:00 — Welcome; Ben sets context for the Part II deep-dive format
00:25 — TJ introduces Southleft and his team's focus on AI + design systems
04:00 — Opening the question: where does your design system live as AI context?
09:07 — Design System Assistant MCP vs. Claude Code-to-Figma: which is better for whom?
10:02 — "Vibe coding" and the emerging pattern of going from code → Figma for UI refinement
15:42 — Community discussion: single source of truth vs. federated systems
15:56 — Eric Steinborn: their source of truth spans JS docs, JSON tokens, Figma, a reference site, and Storybook — and the consolidation effort underway
19:57 — TJ's agentic feedback loop: docs → MCP → code generation → screenshot → validation → iteration
22:56 — Ismail Hamila's AI audit agent: agnostic formats, skills, and checking correct variable intent (not just correct variable usage)
31:18 — Orchestration layers, RAG, and vector databases as an alternative to forcing a single source of truth
31:45 — Ismail's cautionary tale: burning $10 of tokens on a poorly-architected first agent run
34:04 — FigJam spotlight: NY State team's pattern engine; Jennie Yip's design system as AI infrastructure diagram
44:12 — Where does Storybook fit? TJ makes the case for Storybook MCP (via Chromatic) depending on your team
45:35 — Jennie Yip: how packaging everything into an MCP server eliminated AI hallucination
48:04 — Kevin Muldoon in chat: "The blueprint is not derived from the building. Authority flows from origin, not from output."
50:46 — Wrap-up and gratitude for FigJam participation
54:04 — Ben's closing: raw data, FigJam, and coaching resources at https://bencallahan.com

Where to Find the Hosts
Ben Callahan is the founder of Sparkbox (https://sparkbox.com) and the Redwoods Design System Community (bencallahan.com/redwoods), and host of The Question (https://bencallahan.com/the-question). Find him on LinkedIn (https://bit.ly/3T6rd5S).

TJ Pitre is the founder of Southleft (https://southleft.com/). Find TJ on LinkedIn (https://bit.ly/4rsXOBf).

Get the Raw Data
Access the survey data for this episode here: https://bit.ly/4apfR5v

Review the FigJam Notes
Community notes from the deep dive: https://bit.ly/4c9cvFp

Join the Conversation
Subscribe to The Question and join the Redwoods community at https://bencallahan.com/thequestion.