The AI Career Coach That Never Launched

Stella was supposed to be the future of career coaching. An AI chatbot that could guide people through career transitions with the patience of a therapist and the precision of a career counselor, available 24/7 for a fraction of the cost. I built her for a startup called Substep, and she worked beautifully in demos. She never shipped. Here’s what I learned from building a product that didn’t make it.

The Idea

Substep’s thesis was straightforward: career coaching is expensive and inaccessible. A single session with a career coach costs $150-300. Most people who need career guidance — especially those in career transitions or early in their careers — can’t afford it. An AI-powered career coach could provide 80% of the value at 5% of the cost.

Stella was designed to be that coach. She would start by understanding your background, skills, and goals through a conversational intake process. Then she’d help you explore career paths, identify skill gaps, create action plans, prepare for interviews, and navigate the emotional terrain of career change. She wasn’t supposed to replace human coaches entirely — she was supposed to be the accessible first step that helped people figure out whether they even needed a human coach.

Building the Personality

The hardest part of building Stella wasn’t the technology. It was the personality. A career coach needs to be warm but honest, supportive but not sycophantic, direct but not harsh. Getting an AI to hit that balance consistently across thousands of conversation turns is an extraordinarily difficult design problem.

I built Stella’s personality through layers of system prompts, few-shot examples, and guardrails. The base prompt established her voice: empathetic, professional, slightly informal (she used contractions and occasionally humor). The few-shot examples showed her how to handle difficult moments — when someone was frustrated, when someone had unrealistic expectations, when someone clearly needed to hear something they didn’t want to hear.

The guardrails were critical. Stella needed to know when to refer someone to a human. If someone expressed suicidal ideation, she needed to respond with crisis resources, not career advice. If someone described a hostile work environment, she needed to suggest HR or legal options, not just “have you thought about updating your resume?” If someone was clearly dealing with clinical depression that was affecting their career decisions, she needed to gently suggest therapy before career coaching.

Getting these guardrails right took months of testing with real conversation scenarios. Every time I thought I had covered all the edge cases, a tester would find a new one. Someone would describe a toxic workplace in language just vague enough that Stella would miss the signals. Someone would express hopelessness about their career in terms that didn’t trigger the crisis detection but absolutely should have.

The Conversation Architecture

Stella’s conversation flow was structured but not scripted. She had a framework for career exploration that followed a loose sequence: understand the present situation, explore values and interests, identify strengths and gaps, research possibilities, create an action plan. But within each phase, the conversation was freeform — she’d follow the user’s lead and circle back to missed topics naturally.

The trick was maintaining context across long conversations. A career coaching engagement isn’t a single chat — it’s a series of conversations over weeks or months. Stella needed to remember that three weeks ago you mentioned wanting to move to Portland, that last week you were worried about leaving your aging parents, and that today you seem more optimistic than usual. That longitudinal awareness is what separates a coaching relationship from a one-off chatbot interaction.

I implemented this with a structured memory system: each conversation was summarized and tagged with key themes, emotional states, decisions made, and action items. Before each new conversation, Stella would review the summaries to reconstruct context. It worked surprisingly well — testers regularly commented that Stella remembered things they’d mentioned weeks earlier, which made the interaction feel genuinely personal.

Why It Didn’t Launch

Stella didn’t fail because of the technology. She failed because of the business. Substep ran into the same wall that kills many AI startups: the gap between a working demo and a viable product is wider than it looks.

The demo was great. In a controlled setting, with a prepared scenario, Stella was impressive. She asked thoughtful questions, gave useful advice, and felt genuinely supportive. Investors loved it. Beta testers loved it. The problem was scale.

At scale, the edge cases multiplied exponentially. Not just the safety edge cases I’d already worked on, but the quality edge cases. Stella was great with career changers in white-collar fields (the scenario we’d designed for), but she struggled with blue-collar workers, people re-entering the workforce after caregiving, immigrants navigating credential recognition, and veterans translating military experience to civilian roles. Each of these populations needed different domain knowledge, different cultural sensitivity, and different communication patterns.

The cost of getting Stella to handle all these populations well was higher than Substep could fund. And a career coach that only works for white-collar professionals isn’t solving the accessibility problem it was designed to solve — those are the people who can already afford human coaches.

The startup pivoted, then pivoted again, and eventually wound down. Stella never shipped.

What I Learned

More from Stella than from any project that actually launched. Five things, specifically:

Personality design is product design. For a conversational AI product, the personality isn’t a feature — it’s the product. Users don’t evaluate Stella based on her underlying technology. They evaluate her based on how she makes them feel. Getting that right is harder than getting the algorithms right.

Edge cases are the product. The happy path is easy. A career coach that gives good advice to an articulate, self-aware professional with clear goals is not hard to build. A career coach that helps a confused, anxious, inarticulate person figure out what they even want — that’s the hard problem, and it’s the problem that matters.

Safety isn’t a feature you add later. It’s an architecture you build from the start. Stella’s guardrails needed to be deeply integrated into her conversation logic, not bolted on as a filter layer. Every response needed to be evaluated against safety criteria before it was sent, and the criteria needed to be nuanced enough to handle ambiguity.

Demo-ready is not product-ready. The distance between a convincing demo and a shippable product, especially for conversational AI, is enormous. Demos are performed in controlled conditions with prepared scenarios. Products encounter the infinite weirdness of real human beings with real problems that don’t fit neatly into your design assumptions.

Not launching is fine. Stella made me a dramatically better AI builder. The conversational design skills, the safety engineering, the personality architecture — I use all of it in other projects. The fact that the product didn’t ship doesn’t mean the work was wasted. It means the work was practice.

Somewhere in a git repository, Stella is still waiting. She’s patient like that. Career coaches usually are.