CHAPTERS
- 0:00 – 1:17
AI as technical co-founder for an AI-native fashion brand
Yana explains the core premise of Yana Bana: using AI wherever it makes sense across design, production, and even engineering. She frames AI not as a helper but as a stand-in for an entire technical team, while noting that some production steps (like pattern-making) are still not fully solved.
- •AI used end-to-end: design, production planning, and software building
- •Codex/ChatGPT positioned as a “technical co-founder” replacement for hiring engineers
- •Design prompts encode garment details like silhouette, fabric behavior, and construction
- •Pattern generation remains a major unsolved bottleneck; humans and AI run in parallel
- 1:17 – 2:48
Show setup + sponsor: Merge infrastructure for production AI
Claire introduces the show’s mission and tees up the episode’s theme: AI unlocking new creative and business possibilities. A sponsor segment highlights the “everything around the model” work needed for production AI (integrations, permissions, routing).
- •Claire introduces Yana as building an AI-native fashion startup
- •Episode emphasis: AI intersects art/creativity with real-world execution
- •Sponsor: Merge as an infrastructure layer for connecting tools and securing agent actions
- •Key production concerns: reliability, permissions, and cost-efficient model routing
- 2:48 – 5:20
Touring the Yana Bana site: AI-generated garments, visuals, and themes
Yana walks through the Yana Bana site to make the concept tangible—garments are designed with AI and presented with AI-generated collateral (images/videos). She also explains the brand’s visual language that intentionally references computing culture (punch cards, “post-keyboard” motifs).
- •Site showcases multiple AI-assisted garment concepts
- •Designs often begin as hand sketches, then get expanded into technical and marketing assets
- •Brand aesthetic intentionally signals “made in the AI era” (computer/punch-card motifs)
- •Example: “post-keyboard” shirt as a narrative about shifting interfaces
- 5:20 – 6:31
Why AI expands creative execution (not just ideation)
Claire and Yana discuss how AI changes what’s feasible—reducing cost and execution barriers that historically limited creative projects. The focus shifts from AI as “content generation” to AI as a way to actually ship real products.
- •AI lowers practical barriers (time, cost, coordination) to bring designs to life
- •Creativity is “unlocked” when execution is no longer the bottleneck
- •Turning an art experiment into a real manufactured garment is the differentiator
- •Goal: step-by-step from idea → design assets → production → commerce
- 6:31 – 8:02
The fashion prompt stack: encoding garment details like a spec
Yana introduces her “prompt stack,” a structured description of how a garment should look and behave. It includes details beyond shape—fabric flow, movement, and even sound—so AI outputs match her intent and can be iterated toward production-ready direction.
- •Prompt stack covers silhouette, proportions/volume, fabric behavior, and construction details
- •Includes dynamic qualities (movement, drape, sound) to avoid stiff/“paper” results
- •Prompts can be used alone, but are often paired with a sketch for tighter control
- •A reusable core prompt becomes the foundation for consistent creative output
- 8:02 – 10:02
Live demo: generating a jacket and fashion collateral from prompts
Yana demonstrates prompt-to-image generation in ChatGPT, producing a tailored jacket concept and then transforming it into runway-style imagery. She describes her workflow of creating multiple “views” (runway, influencer, product shot) and iterating on details like color, length, and fabric.
- •Prompt produces an initial “product photo” style concept image
- •Follow-up prompts generate editorial collateral (e.g., Vogue runway photo)
- •Workflow emphasizes iteration: colorways, proportions, fabric swaps, and refinements
- •Voice input is central to speed—typing is secondary in her day-to-day flow
- 10:02 – 11:43
Why ChatGPT Images 2.0: better sketch adherence vs. realism tradeoffs
Yana explains why she prefers ChatGPT’s Image Gen (notably 1.5/2.0) over other models: it follows visual directions and sketches more faithfully. The downside is it can overfit to the sketch and lose fabric realism, requiring prompts that explicitly enforce drape and material behavior.
- •Primary selection criterion: faithful compliance with sketches and visual constraints
- •Other models may look “beautiful” but drift toward generic runway similarity
- •Over-precision can make outputs look like paper; prompts must correct for fabric flow
- •Fashion needs novelty and specificity—not nearest-neighbor resemblance
- 11:43 – 13:07
“Prompt as spec”: product thinking applied to creative work
Claire generalizes Yana’s approach into a broader principle: prompts function like product specs/PRDs. Defining “what good looks like” (for garments, photos, or code) improves both AI and human execution and forces useful precision.
- •Prompts work best when treated as a clear spec, not a vague request
- •Defining completion criteria yields more predictable outputs
- •The same clarity benefits humans (“what’s good for AI is good for humans”)
- •Product-style rigor (constraints, requirements) is a creative advantage
- 13:07 – 14:22
Sponsor break + transition: Jira’s Teamwork Graph for agent context
A sponsor segment frames a common agent failure mode: lack of context across tickets, specs, and decisions. It positions Jira’s Teamwork Graph as a way to feed integrated context to coding agents, then returns to Yana’s next steps after image creation.
- •Sponsor: context is the bottleneck for AI coding agents, not code generation
- •Teamwork Graph pulls from Jira/Confluence/GitHub to ground agents in decisions/specs
- •Claimed benefits: higher accuracy and lower token usage
- •Transition back to Yana’s workflow after producing initial visuals
- 14:22 – 16:54
From concept to production path: iteration + Ruth Asawa-inspired dress
Yana shows that not all designs begin with sketches—some start from a moment of inspiration and a single phone prompt. She highlights the friction of safety filters (nudity misunderstandings) and how she adapts prompts (e.g., specifying undergarments) to keep workflows moving.
- •Iteration continues after first images: adjust fabric, color, and construction direction
- •Example: dress inspired by Ruth Asawa’s looped wire sculptures (prompt-first design)
- •Safety constraints can derail fashion imagery; prompts must be carefully framed
- •Production approach depends on garment type (draping vs. patterning vs. fabrication)
- 16:54 – 20:49
Codex + computer use to build CAD and 3D-printable parts (and CLO workflows)
For previously “unmanufacturable” garments, Yana uses Codex with computer use to operate specialized 3D software and generate exportable files like STLs. She emphasizes that agents often perform better inside purpose-built SaaS tools (e.g., CLO) than trying to directly output perfect CAD code from scratch.
- •Workflow: concept → sketches/illustrations → CAD modeling → STL export for 3D printing
- •Codex alone struggles to generate final CAD artifacts reliably; computer use bridges the gap
- •Agents operating real software can outperform direct “generate the file” prompting
- •CLO is used for fashion-specific simulation: patterns + 3D fit on a model
- 20:49 – 24:08
Vendor research and outreach: deep research + drafted emails + Superhuman automation
Yana uses ChatGPT as a research assistant to find manufacturers that match her constraints (e.g., US-based custom apparel). She then has AI draft outreach emails and uses browser/computer use to prepare messages in Superhuman—while keeping herself as the final “send” gate for quality control.
- •Deep research to compile vendor shortlists aligned to specific manufacturing needs
- •AI drafts outreach emails and structures follow-ups
- •Browser/computer use automates tedious CRM-like steps inside Superhuman
- •Human-in-the-loop control: Yana reviews and clicks send herself
- 24:08 – 27:26
Building the full e-commerce stack with Codex: Stripe, databases, and deployment
Yana describes using Codex to build and modify the actual commerce site—adding Stripe preorders, setting up databases for garment voting, and deploying via GitHub/Vercel. She highlights how fast iteration becomes when AI can navigate repos, projects, and configuration changes on demand.
- •Codex used to implement real purchase flows (Stripe) quickly
- •Database tracks community votes to decide what to manufacture next
- •Earlier need for dashboards is replaced by conversational querying via voice
- •Codex can locate Vercel projects and GitHub repos to apply changes end-to-end
- 27:26 – 31:17
Recap + what’s still hard: realism, uniqueness, and pattern generation
Claire summarizes the end-to-end “AI-native business” arc—from creative generation to manufacturing to software ops. Yana identifies the toughest remaining gaps: producing images that are both unique and realistic, and generating accurate sewing patterns from designs.
- •End-to-end pipeline: ideation → visuals/videos → sourcing → manufacturing → e-commerce ops
- •AI enables “previously infeasible” products by collapsing coordination and technical labor
- •Hard problem #1: unique outputs that still look physically realistic
- •Hard problem #2: reliable pattern creation; humans and AI compete in parallel
- 31:17 – 32:52
When AI pushes back: insisting politely for better results
They discuss practical prompting behavior when tools refuse or resist. Yana’s strategy is direct but polite—she believes professionalism improves outputs because it steers the model toward higher-quality communication patterns.
- •Common dynamic: tools initially refuse, then comply with clearer insistence
- •Yana uses a “nice mom” style: polite, direct, and firm
- •Rationale: tone influences the model’s style/quality via training-data priors
- •Politeness is framed as practical optimization, not fear of “AI overlords”
