"Tell me about your AI experience" is four different screens sharing seven words. So is "How have you been thinking about AI in your work?" and "What's your experience with AI products?" and "Have you designed anything with AI?" The surface varies but the diagnostic problem doesn't: commit to the wrong construct and you spend your best material on a screen the interviewer was not running.
This is a diagnostic for reading which construct is live before you respond. It starts where pre-conversation research ends: you are in the room, the question has landed, and you have about five seconds before silence becomes awkward.
Where your AI credibility stands right now
Some of these constructs are perception gaps — you have the evidence, it needs correct routing. Others are genuine evidence gaps where framing cannot close the distance. The distinction governs your response: perception gaps respond to better positioning, evidence gaps respond to honesty and rerouting.
Current state: the Labs (Carrier IQ, Brand Pulse, Retail Velocity) are live as standalone applications, but case-page routes from junochen.com are broken or redirect to login. The Trust essay is publicly accessible. No forward-looking design artifacts have published to the site yet.
Per-construct classification:
- Interaction judgment — hybrid, tilting toward perception gap. Carrier IQ and Brand Pulse are inspectable. The Trust essay is published. Homepage routing to evidence is broken — that part is a perception problem. But neither Lab shows a correction loop (human feedback changing a subsequent output), and the Trust essay describes the philosophy without proving it operates over time. That missing layer keeps this hybrid.
- Prototype fluency — evidence gap. You built the Labs with AI tools, but no published artifact lets a reviewer verify that independently. Verbal description carries the full weight today. When forward-looking design artifacts publish, this shifts to perception gap. That is the single construct whose classification changes with the next site update.
- Production-code depth — experience boundary. TinyFish provided technical grounding, past tense. That is not a production-code shipping record.
- Strategic vision — perception gap. The Trust essay is published, specific, and directly answers this construct.
Four evidence layers
You have four distinct evidence sources. The diagnostic below tells you which to reach for.
- Trust essay — published at junochen.com. The five handoffs, Watch-Verify-Delegate. Your systems-philosophy layer.
- Agentic Labs — Carrier IQ, Brand Pulse, Retail Velocity. Live standalone applications with inspectable interaction surfaces. Your interaction-proof layer.
- TinyFish technical grounding — past tense only. How agent infrastructure works, how retrieval and orchestration function. Your technical-fluency layer, never portfolio proof.
- Forward-looking design artifacts — not yet published. When live, these become your prototype-fluency layer.
The four constructs
1. Interaction judgment
What they actually want to know: how you design the human side of AI products.
Surface forms: "Walk me through how you've designed AI experiences" / "How do you approach trust in AI products?" / "What AI features have you designed?"
Signals that confirm this construct (high confidence): The interviewer is a design leader, head of product, or someone who has shipped AI features. They say craft, patterns, interaction model, user trust, transparency. The follow-up asks about a specific design decision — "Walk me through how you decided what to show the user when the agent is working."
Signals that suggest it (moderate confidence): The role posting emphasized "design for AI-powered experiences" or "human-AI interaction" without mentioning code, architecture, or technical implementation.
Clarifying move: "I've designed oversight and trust patterns for agentic systems — would it be most useful to walk through the interaction model for a specific one, or talk about the principles I've developed across several?" This confirms the construct and lets the interviewer choose altitude.
Lead with Carrier IQ. The Approve Bind gate, the stage-visibility model (Session, Navigate, Fill, Extract, Verify), and the operator notes field are your strongest inspectable proof of designing human oversight into an agent workflow. If they want breadth, bridge to Brand Pulse — source-labeled evidence cards, the editable mission field, cancel/rerun/restore controls. Different expression of the same judgment: making agent reasoning visible and interruptible.
Where the evidence stops: Neither Lab shows a correction loop — a sequence where human feedback changed the system's subsequent output. The Trust essay's Watch-Verify-Delegate framework describes the philosophy but you cannot point to an inspectable record of it operating over time. If the follow-up goes there, name the framework, describe the design intent, and say the artifact is in development. Do not stretch Carrier IQ's Approve Bind into a learning loop. It is a gate. The interviewer screening for interaction judgment will know the difference.
2. Prototype fluency
What they actually want to know: can you build with AI tools, independently, without waiting for engineering.
Surface forms: "How do you use AI in your design process?" / "Do you prototype with AI tools?" / "How technical are you day to day?"
Signals that confirm this construct (high confidence): The interviewer asks about your workflow or process, not your product work. They name specific tools — Claude Code, Cursor, Figma AI. Docebo's published interview process makes this screen explicit: Step 3 evaluates "how you use AI tools as a multiplier in your design process." Headway's posting frames code prototyping with AI tools as core to the role.
Signals that suggest it (moderate confidence): The interviewer is an engineering manager or technical PM. The question feels like it is really asking: Will you need an engineer to build every prototype, or can you move on your own?
Clarifying move: "Are you asking about how I use AI tools in my design process, or about the AI-product work I've shipped?" This is the most important disambiguation in the set. Prototype fluency and interaction judgment sound adjacent but require completely different evidence. Getting this wrong means spending five minutes on Carrier IQ's trust architecture when the interviewer wanted to hear about how you built it.
Lead with describing how you built the Labs — the tools, the iteration speed, the specific moments where a working prototype revealed something a static mockup would have missed. This is verbal evidence today. It is credible when specific (name the tool, describe the decision point) and weak when general.
Where the evidence stops: No published forward-looking artifact a reviewer can visit. When those artifacts publish, this construct shifts from evidence gap to perception gap. Until then, verbal description carries the full weight. Keep it concrete and brief — specificity carries more weight than volume.
3. Production-code depth
What they actually want to know: have you shipped AI systems into production.
Surface forms: "Have you shipped AI products?" / "What's your experience with ML systems?" / "How close to the code do you work?"
Signals that confirm this construct (high confidence): The interviewer is an engineer, a technical co-founder, or someone whose background is in ML/AI engineering. They ask about architecture, models, evaluation, production, or shipping code. OpenAI's Codex posting states this directly: candidates should "bring a strong technical background and have recently shipped code to production."
Signals that suggest it (moderate confidence): The follow-up after your initial answer shifts to implementation specifics — not "how did you decide what to show the user" but "how does the system handle failure states at the infrastructure level" or "what's your experience with model evaluation."
Clarifying move: There is no clarifying move that helps here. Clarification is useful when you have evidence for the right construct but might present evidence for the wrong one. When the construct is production-code authorship and you do not have a recent production-code record, clarification delays the gap becoming visible without closing it.
What to do instead: TinyFish gave you technical grounding — you understand how agent infrastructure works, how retrieval and orchestration function, what the engineering constraints are on the design decisions you make. Use that grounding in past tense, as context for your design judgment, to show you are not handing off wireframes without understanding what happens downstream. But do not position it as equivalent to a production-code shipping record. The interviewer screening for this construct will see through that immediately.
For roles where this is the primary screen — OpenAI Codex is the clearest example — this is an experience boundary. Better framing can close a perception gap. Building artifacts can close an evidence gap. An experience boundary is different: it means rerouting to a different entry point. If the diagnostic confirms production-code depth is the primary screen, that is useful qualification intelligence about the role, not a positioning problem to solve.
4. Strategic vision
What they actually want to know: where is AI going and what does that mean for design.
Surface forms: "How do you think AI changes [product category]?" / "What's your vision for design in an AI-first company?" / "Where do you see human-AI interaction heading?"
Signals that confirm this construct (high confidence): The interviewer is a CPO, VP Product, founder, or senior executive. They lean back. The question feels expansive — they want a point of view, not a case study. Follow-ups are future-oriented: "How does AI change [the product category]?" or "What does design leadership look like in an AI-first company?"
Signals that suggest it (moderate confidence): The role is Director+ or the posting emphasizes "shaping the vision" or "defining the role of design" in AI product development. Stripe's posting asks candidates to "move fluidly between big-picture thinking and interaction-level craft" — both constructs may be live in the same conversation.
Signal to watch (use with caution): A senior executive opens with a future-oriented question but their background is technical. This can look like a vision screen but function as a warm-up before they drill into production-code depth or interaction specifics. If the interviewer's LinkedIn shows an engineering or ML background, treat the expansive opening as provisional. Answer the vision question, but watch the follow-up before committing to this construct for the rest of the conversation.
Clarifying move: "I have a specific point of view on where human-AI interaction is heading — want me to start there, or ground it in a system I've built?" Let the interviewer choose altitude. If they want the thesis, lead with the Trust essay's framework — the five handoffs, the progression from Watch to Verify to Delegate. If they want it grounded, start with Carrier IQ as a testable expression of the thesis and pull up to the principle.
Lead with the Trust essay. It is published, it is specific, and it demonstrates exactly what is being screened: a forecast about where AI interfaces are going, articulated clearly enough that someone can disagree with it. The Labs then serve as proof you can turn a thesis into something testable.
Where the evidence stops: Vision without production evidence can read as conceptual. If the follow-up shifts from "where is this going" to "show me something you built that proves it," you have moved into construct 1 or 2. Re-route accordingly.
After you answer
The follow-up question tells you whether you read the construct correctly. If the next question goes deeper on the same thread, you read it right. If it pivots — you answered about interaction judgment and they follow up about code — you misread, and you have one clean chance to re-route before the evaluative frame sets. Watch for the pivot. It is the most reliable real-time signal you will get.
At 8:30 AM
The failure mode is answering for the wrong version of "AI experience" and spending your strongest material on a construct the interviewer was not screening for. Five seconds of listening before you speak will tell you more than anything you rehearse beforehand.
- Docebo's published interview stages: The only one of the four target employers that maps AI constructs to specific interview moments, with Step 3 explicitly screening for AI-tool fluency as a workflow multiplier rather than production-code authorship.
- Atlassian's IC-versus-manager bifurcation: Their public design interview handbook separates product thinking and craft excellence into distinct 30-minute stages, illustrating how constructs that feel blended in conversation may be assessed independently.
- The correction-loop artifact gap: None of the three approved Labs shows human feedback changing a subsequent system output, which the KB identifies as the single highest-value missing proof for closing the interaction-judgment construct from hybrid to perception gap.
- Category-spanning selection costs: A study of 964,034 Elance bids found that employers penalized erratic category movement more than incremental shifts, which has implications for how the Head-of-Product-to-design-leadership sequence reads to evaluators screening for AI-design commitment.