Before you write outreach or pick which proof to attach, settle one thing: what type of AI-design capability is this role filtering for?
Call that type a construct. Five of them operate as genuinely separate screens in current postings:
- Interaction judgment and trust architecture
- Code-assisted prototyping fluency
- Production-code authorship
- Model-behavior design
- Agent-workflow orchestration
The vocabulary overlaps badly across all five. "Agentic" currently appears in postings screening for at least three of them. The word by itself tells you nothing. The construct tells you which story to lead with and which proof to send.
Below are three signal types you can read before you enter any conversation: the patterns that point to each construct, and the organizational fear driving the language. The fear is the part worth holding onto, because it lets you read a posting whose vocabulary you have never seen before.
Posting Language
The fastest read. Scan the phrases, but weight the fear behind them more heavily.
"Trust," "policies," "edge cases," "sensitive moments," "review states," "approval," "recovery," "source visibility"
Construct: Interaction judgment and trust architecture. The fear: users act on AI output without enough evidence, enough control, or a path back. Stripe's Data & AI posting is the cleanest current example: ambiguous product direction, complex systems and policies, sensitive moments. Docebo adds source visibility, approval loops, privacy safeguards.
"Claude Code," "Cursor," "prototype in code," "working prototypes," "pressure-test before engineering commits"
Construct: Code-assisted prototyping fluency. The fear: design stays abstract and slow, and every idea needs an engineer to translate it. Headway names Claude Code and Cursor as core design tools and expects designers to ship improvements to the codebase. The tools are named because the company wants evidence you already work this way. Willingness to learn is not what is being bought.
"Shipped code to production," "PRs every week," "production codebase," "owns prompts/code/data"
Construct: Production-code authorship. The fear: a designer who stops at the prototype and needs engineering for every technical change. OpenAI Codex asks for a strong technical background and recent production-code shipping. The Browser Company expects multiple pull requests a week. When "shipped code to production" sits in requirements rather than preferred skills, treat it as a qualification boundary: the evidence exists or it does not, and no reframe closes the gap.
"Model as product," "design model behavior," "data strategy," "evals," "regressions," "model selection"
Construct: Model-behavior design. The fear: the behavior they want cannot be specified, measured, or held steady across releases. OpenAI's Model Designer role treats the model itself as the design surface: predicting and designing model behavior, devising data-collection strategies, understanding how changes in data move output. That is ownership of what the model does, not of the interface around it.
"Multi-agent," "orchestration," "roles," "tool access," "shared context," "handoffs," "platform primitives"
Construct: Agent-workflow orchestration. The fear: multiple actors, agents and people, coordinating incoherently even when each individual capability works fine. Mural is the clearest current example: agent behavior, communication, handoffs, shared context, platform primitives for multi-agent collaboration. Mural also draws an explicit line between code as a design exploration tool and ownership of production code, which is where orchestration separates from production-code authorship.
Headway uses it while screening for code-assisted prototyping. Docebo lists it as a bonus while screening for trust architecture. Mural uses it while screening for orchestration. The adjective is not the screen. Look at what the role is accountable for delivering.
Team Composition and Reporting Structure
Reporting lines and peer roles tell you which organizational problem the hire is meant to solve. Where the posting language stays ambiguous, structure often resolves it. Sometimes it does not, and I have flagged where.
Role housed in Engineering AI, reporting to a Model Behavior Lead. Construct: Model-behavior design or production-code authorship. The designer is expected to touch prompts, data, evals, and model behavior directly. Which of the two applies depends on whether accountability lands on behavioral quality or on the technical product surface. The Browser Company's role sits here.
Role embedded with researchers and engineers on a frontier technical product. Construct: Production-code authorship when the posting asks the designer to ship code personally. Model-behavior design when it centers behavioral judgment, data strategy, and writing quality instead. Check whether the posting asks for technical collaboration or for the designer to ship production code themselves. OpenAI Codex crosses that line. OpenAI's Model Designer does not.
Role inside Product Design with PM, engineering, and research peers. Construct: Does not resolve through structure alone. The most common configuration and the least diagnostic. Headway and Stripe both use it. Headway resolves toward prototyping through named tools and codebase participation; Stripe resolves toward trust architecture through policies, systems thinking, interaction craft. Structure will not get you there. Go back to the language.
Role reporting to a design executive inside a regulated vertical. Construct: Interaction judgment and trust architecture. UiPath places a Principal designer under a Senior Director of Design for Vertical Solutions. The named design problem, a human confirming or overriding an AI claim decision, points to trust architecture. Code-assisted prototyping appears here as working method, not as the capability being tested.
Design and Research team asked to create platform primitives for agent coordination. Construct: Agent-workflow orchestration. Mural asks the designer to define agent roles, delegation, tool access, handoffs, shared context, and coordination primitives across the system. The separation between code as exploration tool and code as owned production surface is the structural marker. A design team building platform-level agent architecture without owning production code is an orchestration screen.
Founder-level reporting in a very small AI-native company. Construct: Broad AI interaction design. Sauna is fifteen people, seed stage, designer reporting to the CEO. The role covers interaction architecture, design systems, research, and prototyping against real models. Code proficiency is advantageous, not mandatory. Expect a broad screen rather than a specialized one.
Code-assisted prototyping fluency has no distinctive structural signature. It shows up inside Product Design teams, under design executives in regulated verticals, and inside Product organizations. For this construct, posting language is your primary diagnostic. Named tools and codebase participation expectations answer the question regardless of where the role sits.
Company Profile
Stage and product type create tendencies. They do not determine constructs. OpenAI is currently advertising a model-behavior role and a production-code-authorship role at the same time. Treat profile as a prior, and let posting language and team structure confirm or override it.
| Profile pattern | Likely construct | Organizational concern |
|---|---|---|
| Frontier model company, model-design function | Model-behavior design | The model is the product; the designer changes what it does |
| Technical AI product team building a coding/dev tool | Production-code authorship | The designer operates inside the technical work, not around its interface |
| Late-stage vertical platform integrating AI into existing workflows | Interaction judgment and trust architecture | AI decisions affect consequential outcomes; humans need evidence, control, recovery |
| Established collaboration platform adding an agent layer | Agent-workflow orchestration | Coordination across multiple actors, not a single AI surface |
| Growth-stage platform making AI part of design practice | Code-assisted prototyping fluency | The company wants designers who already build in code |
| Very small founder-led AI-native startup | Broad interaction design + real-model prototyping | One designer covers everything; the screen is for original interaction thinking |
Your Evidence by Construct
One-line pointers. The construct-specific dossiers own the depth.
Interaction judgment and trust architecture. Lead with Carrier IQ for review states, evidence attachment, and the human authorization gate, plus the Trust Is the New Interface essay. For enterprise-scale trust, pair with Thermo Fisher mySupply, completed as Product Design Director at BCG Digital Ventures.
Code-assisted prototyping fluency. Lead with Brand Pulse or Retail Velocity as working applications. Retail Velocity carried a TinyFish infrastructure dependency; disclose it, and do not present it as proof you built the underlying agent stack. Neither Lab is currently linked from your homepage, so send the direct URLs.
Production-code authorship. You do not have a public repository, commit history, or record of shipping production code. This is a qualification boundary. The Labs show design judgment expressed through working software; they do not substitute for what this construct asks to see.
Model-behavior design. The Trust essay and Brand Pulse demonstrate behavioral judgment and visible evidence states. They do not establish ownership of training data, prompt systems, eval pipelines, or regression gates.
Agent-workflow orchestration. Brand Pulse shows parallel agent states and evidence aggregation. Carrier IQ shows staged agent work ending in human authorization. Neither proves a published platform architecture for multi-agent memory, tool allocation, or cross-agent handoffs. The Delegation Contract would be the strongest proof here if it existed. It remains specified and unpublished.
Before You Send Anything
Identify the construct. Check what you can actually send against it. If the construct asks for proof you do not have, that is a qualification boundary, and the time to find that out is before you write, not after the second interview.
- OpenAI screens diverge internally: The same company advertises model-behavior design and production-code authorship as separate roles with different requirements, which means company name alone cannot tell you which construct you are walking into.
- Mural's code boundary language: Mural's Lead Product Designer posting explicitly separates code-as-exploration from production-code ownership, making it the clearest current example of how an orchestration role draws that line.
- Labs still missing from homepage: Brand Pulse, Retail Velocity, and Carrier IQ remain reachable at their standalone URLs but absent from the current homepage routing, which means every construct-specific outreach requires you to send direct links rather than relying on a reviewer's natural navigation.
- Delegation Contract remains unbuilt: The artifact specified to close the agent-workflow orchestration gap — showing permission, action, correction, and residual state — is still in the design-context backlog with no public URL, leaving that construct the least evidenced in your current portfolio.