Nine Act-tier sheets, sorted by the layer at which each evaluator is competent to judge you.
| Company | Role | Altitude | Open with |
|---|---|---|---|
| Stripe | Sr Staff PD, Agentic Commerce | Craft | Alibaba decision surfaces |
| Abridge | Staff PD | Craft | Red Cross eligibility screen |
| Ambience | Staff PD | Craft | Thermo Fisher three-reader record |
| Spring Health | Staff PD, AI Interaction | Craft | Allē two-sided visibility |
| Maven | Sr Staff PD, Care Delivery | System | Carrier IQ, live |
| HackerOne | Director, AI Product Design | System | Five control-transfer points |
| Altana | Head of Product Design | System | Red Cross one model, three views |
| Vanta | Head of Design | Leadership | Thermo Fisher failure-mode modules |
| Amplitude | Head of Product Design | Leadership | Equinox+ five brands, 90 days |
The diagnostic, in thirty seconds
Every posting is a sentence about accountability. Locate the clause naming what you will own. Then look at what the verb takes as its object. The object is the altitude.
Craft altitude. The objects are surfaces and states: screens, flows, motion, fidelity, end-to-end execution. This reader wants the iteration sequence. Show what shipped and the three versions that didn't, and say what killed each one.
System altitude. The objects are patterns and primitives: reusability, grammars, handoffs, review queues, a design system composed by something that isn't a person. This reader wants the model. Show the abstraction and what it made possible downstream.
Leadership altitude. The objects are people and standards: headcount, definitions of done, launch review, bar-holding, growth. This reader wants a criterion that survived your absence from the room.
Two rules make the read portable.
When two altitudes show up in one posting, the accountability clause wins. "You will be responsible for" beats "experience with." Qualifications sections run inflated because HR wrote them. Ownership language comes from the person you'd report to.
Altitude is not seniority. Abridge posts a Staff IC role and inspects at craft. Altana posts a Head of Design and inspects at system. Title sets the compensation band. The object of the accountability verb sets which portfolio layer gets opened.
Thermo Fisher appears twice below. So does Red Cross. Same projects, four different cuts. A case is material. The buyer determines the cut.
Surfaces and states. Patterns and primitives. People and standards.
Before you send anything
Checked August 1: every full case page on junochen.com bounces an unauthenticated visitor to a password prompt. The teasers are public and they are substantive. Alibaba, Thermo Fisher, Red Cross, Equinox+, Allē. Link teasers in cold outreach. Never link a full case. From a stranger, a password wall reads as carelessness before it reads as confidentiality.
Correcting the Design Context File, which on July 24 listed Carrier IQ as public and clean. Half of that held. The working app opens without a password. Its write-up page doesn't. Send the app.
Every number below is a portfolio-owner claim, not an audited figure. Say that plainly if a recruiter presses one. TinyFish is context, not evidence: it establishes that you build agent systems now, and it supplies no case and no metric.
Craft altitude
Stripe — Senior Staff Product Designer, Agentic Commerce
Problem. My inference, not their words. The mandate spans merchants, agents, consumers, LLM conversations, checkout, and Link. One checkout has to resolve correctly for a human buyer, a merchant integration, and a model transacting on someone's behalf, without fragmenting into three interfaces.
Open with.
"At Alibaba I rebuilt the components where a B2B buyer decides to proceed: the search result and product detail surfaces where a trust signal either closes the credibility gap or the buyer leaves. Across 25M+ desktop sessions."
Number. $50B+ GMV, on a marketplace where the buyer was committing to a high-value order with a supplier they had never met.
Only you. This slate divides into candidates with consumer checkout craft and candidates with agent-system literacy. You hold both: the commerce decision surface at marketplace scale, and agent-run web workflows you build daily at TinyFish. The role sits on that seam.
Don't lead with. Outcomes. A finished screen with a metric under it reads as someone who has stopped touching the work. Bring states, rejected versions, and the interaction you argued about.
The clinical buyer is four buyers
Every recruiter groups Abridge, Ambience, Spring Health, and Maven. So does every candidate. Don't. They share one condition and diverge on everything that determines your opening line.
The condition: whoever receives the AI output carries professional consequence for what happens next. That pushes the interface requirement past explainability. Understandable at the moment of approval is not sufficient. The artifact has to be defensible later, to a reviewer who wasn't there, sometimes years later, sometimes under regulatory examination.
The axis separating this group from every other AI-native buyer is reversibility. A wrong agentic checkout at Stripe gets refunded. A wrong triage at HackerOne goes back in the queue. A signed clinical note does not get unsigned. What these evaluators are screening for is whether you have designed against a decision nobody can take back.
Two of your cases answer that, on different halves. Thermo Fisher is the regulated-review proof: a record built to survive audit. Red Cross and Allē are the non-expert-under-pressure proof: real stakes, no domain training, deciding fast. Your agentic work supplies the third piece, calibrated autonomy, the machine doing part of the job and stopping cleanly.
Then they split. Note also that clinical is a buyer cluster, not an altitude. Three of the four inspect at craft.
- Clinical-record review. Abridge and Ambience. A clinician reviews generated documentation and signs it into the patient record. Ambience runs materially broader than a scribe: clinicians review drafts carrying suggested ICD-10 and CPT codes, proposed medication orders queue in Epic for approval, and compliance and revenue-cycle staff read the same artifact afterward.
- Live decision support. Maven. Provider tooling used inside an appointment. Nothing gets signed. Uncertainty communicates in real time or it doesn't communicate.
- Member-facing with clinical guardrails. Spring Health. The direct reader is frequently the member. Authority enters through routing, escalation, and hard limits on what the agent may say.
Send the wrong half of your proof to the wrong group and you have told them you think healthcare is one market.
Abridge — Staff Product Designer (craft)
Problem. Their language: "transforming complex clinical workflows and AI capabilities into intuitive and delightful user experiences." The product ships Linked Evidence, mapping generated summary text back to source so a provider can verify before trusting.
Open with.
"For the American Red Cross I designed the screen where a caseworker decides, under disaster-response pressure, whether a household qualifies, with eligibility logic and audit trail rendered inside the same view as the approve action."
Confirm the co-location against the case before that sentence goes into a cover letter. The teaser gets you eligibility and audit logic. The claim that they sit in the same view as the approve action is yours to stand behind.
Number. 1,689 cases processed in the first two weeks of national deployment.
Only you. You have designed verification that a rushed person actually performs instead of skipping. Verification fails at the craft layer, not the concept layer, and you have the production failure states that taught you so.
Don't lead with. Anything organizational. This is an IC role judged on pixel, motion, and interaction quality. Org scale and mandate-building read as overqualified, and overqualified here reads as disengaged.
Ambience Healthcare — Staff Product Designer (craft)
Problem. Their phrasing: own and accelerate the end-to-end experience across Clinical AI products. One reviewed artifact carries documentation, coding, compliance, and orders at the same time.
Open with.
"At Thermo Fisher I designed a supply-chain release workflow across six pharmaceutical partners where a single record had to satisfy three readers with different accountability: the operator releasing it, the QA reviewer signing it, and the auditor reconstructing the decision later under regulatory constraint."
Number. 100% partner adoption across six pharmaceutical partners and nine sites. Every partner on the platform, none of them a captive user of one internal team's tool. If you can't confirm adoption was voluntary, cut everything after "nine sites."
Only you. You design for the downstream inspector, not only the approver. Nearly every AI-documentation candidate optimizes the signing moment. Ambience's coding and compliance surfaces get read by people who arrive after the clinician has moved on, and that reader has nobody to ask. If they push toward scalable pattern systems, follow them there. The opening still has to be craft.
Don't lead with. Delight, consumer polish, or Equinox+. This buyer hears aesthetic fluency as evidence you have never worked inside a controlled release process.
Spring Health — Staff Product Designer, AI Interaction Design (craft)
Problem. Verbatim: "Design the human-AI interaction experience for Spring Health's user-facing AI agents across web and mobile platforms." Intent routing, hallucination handling, and language for people in distress.
Open with.
"On Allē I designed a two-sided system where consumers made treatment decisions in an iOS app while the practice worked the same relationship from a back office, and the real design problem was what each side could see, say, and act on without the other in the room."
Number. 47% reactivation, which is the number proving the handoff between the two sides actually held.
Only you. You have built consumer surfaces whose stakes resolve on the professional side of a boundary. Their Guide doesn't diagnose. It decides when to stop and route. That boundary is a design object and you have shipped one.
Don't lead with. Clinician-signature framing. The reader of this agent's output is often the member, and presenting as a clinical-documentation designer misreads who is on the other side of the screen.
Maven Clinic — Senior Staff Product Designer, Care Delivery (system)
Problem. Their own framing: "designing our provider tooling that our virtual care network of clinicians and other providers use during appointments," including how AI communicates uncertainty, defers to human care, and contributes confidence indicators and handoff patterns to the design system.
Open with.
"I built and published a broker-facing agent that runs parallel carrier workflows and returns confidence-scored quotes side by side, where confidence is grounded in data completeness and the division of labor is stated outright: the agent assembles, the person evaluates."
Number. Not a number. A URL. carrier-iq.junochen.com, built and published independent of TinyFish, running without a password, the agent's stages visible while it works.
Only you. Maven wants uncertainty patterns as reusable system components. Most candidates describe them. You can send one that executes and name which handoff points it already instantiates, in-progress visibility in the stage display and output review in the confidence-scored comparison, then name the two you would build next.
Don't lead with. The VP framing from the Issue #5 Maven sheet. This posting is Senior Staff, prototypes in code, grows peers laterally. Different altitude, different first paragraph.
System altitude
HackerOne — Director, AI Product Design
Problem. The posting names agents as the primary workforce while humans retain "oversight, accountability, and strategy," alongside agent-to-agent interaction, human-in-the-loop control, and risk visualization. Someone has to design the control surface between those two facts.
Open with.
"I designed and documented the agent supervision layer as five points where control transfers between person and system, intent-setting, in-progress visibility, output review, decision gate, loop feedback, and then built against it."
Number. Five named transfer points, each mapped to a shipped surface in the portfolio. One of them is already literal: Carrier IQ exposes five discrete agent stages (Session, Navigate, Fill, Extract, Verify) while it runs, which is in-progress visibility built rather than described.
Only you. You wrote the control taxonomy publicly before this posting existed, and it carries a test this evaluator can run against their own product tonight: accuracy is what the model achieves, consistency is what the design delivers. Competing candidates will describe human-in-the-loop as a value.
Don't lead with. Security domain expertise. You don't have it, they know it, and the mandate is oversight architecture rather than vulnerability triage.
Positioning only. Outreach targets, sequencing, and draft messages for HackerOne live in the Opportunity Brief.
Altana — Head of Product Design
Problem. Their words: a design system as "a semantic, machine-readable language" of primitives and interaction grammars that people and agents compose, carrying citations, confidence, consent, review queues, and graceful failure.
Open with.
"For the Red Cross I collapsed six disconnected systems into a single information model and expressed it as three role-specific views, so the same underlying record rendered differently depending on who was accountable for the next action."
Number. Three views from one model, holding identically across a national deployment where urgency routing and eligibility logic could not be allowed to diverge.
Only you. You have built one semantic layer that resolves into multiple interfaces. That is their stated architecture. Most Head of Design candidates have built component libraries, a different discipline wearing the same word.
Don't lead with. Team size or hiring plans. Their problem statement is architectural. Design the grammar before you discuss who speaks it.
Leadership altitude
Vanta — Head of Design
Problem. Their phrasing: "build the systems that hold that bar consistently: design system, definitions of done, and launch review frameworks," across roughly 40 people and five product areas. The verbs are build and establish, not maintain. Read that as absence, and verify it in the screen.
Open with.
"At Thermo Fisher I ran 32 interviews that surfaced five named failure modes, then converted each failure mode into its own module, scoped by the specific breakdown it had to prevent, which is how 'done' stopped being a matter of whose opinion carried the room and became a question of whether the named failure could still occur."
Number. A reported $20M margin opportunity and 83% IRR, on a system whose quality criteria came out of research rather than out of a reviewer's judgment.
Only you. You have derived a definition of done from evidence. Most candidates at this level have enforced standards they inherited. Vanta is asking someone to author them.
Don't lead with. The Trust essay. Issue #5's Vanta piece routed proof through Trust plus Alibaba. Invert it. The handoffs are evidence for an oversight mandate, and this mandate is quality mechanism. Then the authority risk: authoring definitions of done across five product areas with no existing mechanism is a job nobody does without a reporting line and a real launch-review veto. Ask about both.
Amplitude — Head of Product Design
Problem. Verbatim: "AI makes it cheap to ship UI changes, which makes it easy for quality to erode quietly." They want a standard that holds across 15 designers without turning into an approval bottleneck, enforced by someone still in the work.
Open with.
"As Product Design Director on Equinox+ I unified five brands into one platform and took it from zero to MVP in 90 days, which meant defining what consistency meant before speed had a chance to erase it."
Number. 4.8 stars at launch, on a 90-day build. The case where velocity did not quietly cost quality.
Only you. You can describe the mechanism that caught drift, not the intention to prevent it. Their stated fear is quiet degradation. Answer with the detection method.
Don't lead with. Executive framing. The posting explicitly rules out management-only, pure IC, design ops, and approval committee. It also names no reporting line, which is your first question in the screen. Position as the player-coach who pairs on hard UX problems and gives direct, specific feedback, because those are their words and they chose them carefully.
- Autonomy and interruption move together: Anthropic's coding-agent telemetry shows experienced users auto-approving more often and interrupting more often, which breaks the Watch → Verify → Delegate ladder into two independent controls — read the autonomy measurement work before you defend the ladder in a HackerOne or Altana screen.
- Altana's equity window is unverified: The most recent official financing record is a $100M Series B from October 2022, which tells you nothing about current valuation or dilution — treat compensation as base-plus-unknown until you get a grant conversation.
- What "done" means for a probabilistic product: Figma's evaluation process runs a roughly 1,000-example corpus, nightly model comparisons, and a mix of deterministic tests and human judgment, documented in First Round's account — the closest public answer to Amplitude's quality-erosion question.
- Governance is not a UX preference: The NIST AI Risk Management Framework asks for documented go/no-go decisions, override statistics, and post-deployment appeal and recovery mechanisms, which is why Thermo Fisher's QA gate reads as product architecture to a clinical buyer rather than compliance theater.