Designing an AI recruiting product from zero, as the sole product designer. From a search box to a team of sourcing agents, an assistant that coaches the recruiter, and outreach that runs from their own inbox.
Context
Startup recruiting is broken in a specific way: the companies that need great talent the most have the fewest resources to find it. They can't afford agency recruiters. They don't have time to manually source across five different platforms. And the tools that exist are either built for enterprise or built for keywords, not for the nuanced, context-rich way startup hiring actually works.
Wellfound had a decade of startup hiring data. 10M candidates who came to Wellfound specifically for startup roles, 500M+ enriched external profiles, and 13 years of signals: company stages, inferred skills, teammate data. But no product that put all of it to work for a self-serve recruiter. That was the opportunity.
Discovery & Research
Before drawing a single screen, I needed to understand the landscape. Who are startup recruiters? What does their day actually look like? Where do existing tools fail them, and where do they succeed?
The first list is fine. The stall after it is the product.
Affinity mapping from 20+ user interviews, clustering pain points into themes
Strategy & Framing
From the research, four clear jobs emerged. These became the structural backbone of the product, and every feature maps back to one of these.
Sourcing agents that understand context, not keywords. Describe the role in plain English; the agent evaluates company stage, trajectory, inferred skills, and candidate intent, and explains every fit.
AI-personalized sequences sent from the recruiter's own inbox. Auto Pitch runs outreach around the clock; the recruiter can review every message or let it run. Candidates see a person, not a platform.
Building the initial list is the easy part. Ask Wellfound, the assistant inside Reach, suggests the next search when one stalls, expands the pool to adjacent companies, rewrites outreach that isn't converting, and flags layoffs and talent movement.
AI evaluations, standardized resumes, and company context for every applicant. Inbound and agent-sourced candidates reviewed in one shortlist, filterable by which agent found them.
Design Process
As the sole designer on a fast-moving team, I had to move between altitudes constantly. Strategic framing one day, pixel-level details the next. And the product has been redesigned in place more than once as the AI got more capable. Here's how the process worked.
The first design challenge was structural. Recruiting touches searching, evaluating, messaging, tracking, and analyzing, all tightly coupled. I mapped every user flow before opening Figma, working through how a recruiter's day would actually move through the product.
The IA has since reorganized around agents. A project holds its agents, each agent owns a strategy, a sequence, and its own funnel, and the shortlist is where everything lands.
I explored 3–4 directions for each major surface before converging. The candidate profile was the hardest. It needed to show deep context (company history, inferred skills, career trajectory) without overwhelming a recruiter who's evaluating 50 profiles in a session.
Low-fidelity wireframes first, shared early with engineering to catch feasibility issues. Then medium-fidelity explorations tested with 5–8 recruiters per round.
The agent card went through the same treatment. It had to say, in one glance, what the agent is looking for and how it is doing (evaluated, matched, pitched, replied) without turning into a dashboard.
says what it’s looking for ✓
says how it’s doing ✓
not a dashboard ✓
We ran moderated usability tests at every major milestone: search experience, candidate evaluation, outreach flow, and pipeline management. Each round surfaced things we hadn't anticipated.
The biggest pivot came from testing the search experience. Our first version was powerful but complex, with too many filters and too much control. Recruiters told us they wanted to describe what they were looking for in plain English and have the AI figure out the rest. That insight reshaped the entire search paradigm.
The second big pivot came later. Recruiters ran one search, got a good list, and stalled. Testing showed they didn't want more filters, they wanted a colleague: "what should I try next?" That became Ask Wellfound, and the chat moved from a side panel to the home of every project.
01 sentences, not boolean
02 “what do I try next?”
03 a colleague, not a dashboard
Key Design Decisions
A search is a one-shot. An agent keeps working. We reframed the primary object from "a search you run" to "an agent you launch." It has a strategy written in plain English ("senior backend engineer who's worked at a Series B startup and knows distributed systems"), it evaluates candidates continuously, and you can run several at once to test different strategies for the same role. The UI had to make a running agent feel trustworthy: what it is doing right now, what it found, what it did on your behalf.
Our data showed that outreach from a recruiter's actual email gets 50% higher response rates than platform messages. So we designed the entire outreach flow around the recruiter's real identity: their inbox, their voice, their signature. AI personalization fills each message uniquely per candidate, follow-ups are automatic, replies land in Gmail and in the app, and Reach drafts the response from the recruiter's guidelines. Auto Pitch is the recruiter choosing to let it run.
Most recruiting tools show you data. Reach tells you what to do about it. Ask "what other companies like Vimeo can I target?" and Ask Wellfound researches the market, expands the pool, and creates the agent for you. It also meets recruiters where they already work: Slack, Gmail, and any AI assistant through MCP.
Outcomes
Wellfound Reach launched as a self-serve AI recruiting product and has kept shipping: unlimited sourcing agents, Auto Pitch, Ask Wellfound, email and LinkedIn sequences, ATS and AI-assistant integrations. Concept to first ship in under a year, and two major re-architectures since.
Reflection
Being the sole designer on a 0→1 product is a masterclass in prioritization. You can't design everything perfectly, so you have to decide what matters most, ship it, learn from it, and iterate. The hardest part wasn't any individual design problem; it was knowing when to go deep and when to keep moving.
The biggest lesson: AI products need more design thinking, not less. When the technology can do anything, the design challenge shifts from "what's possible" to "what's useful." Every feature in Reach was filtered through a simple question: does this save the recruiter time, or does it just look impressive?
Designing for agents changed the job again. When the product acts on the recruiter's behalf, the design work is less about screens and more about trust: showing what the agent is doing, why it chose someone, and leaving the recruiter a way to steer without having to babysit. The best interface for an autonomous system is one that explains itself.
Working this closely with engineering also changed how I design. I learned to think in system constraints early, to design for edge cases that would have bitten us in production, and to treat the handoff not as a deliverable but as an ongoing conversation.