First Day

You only truly know what you can explain..

Comment

Inspiration

I've managed telecom infrastructure projects for ten years. This year, while revising for a project management certification, I noticed the concepts I actually remember are the ones juniors forced me to explain over the years, float, critical path, scope creep. I don't know them because I read them; I know them because someone asked me a question and my answer had to hold up. It turns out this has a name: the protégé effect, you learn best by teaching.

I've also onboarded a lot of juniors: it takes months, the knowledge transfer is huge, and the week-one questions are always the same. So I built the junior.

What it does

First Day inverts the AI tutor. GPT-5.6 plays a new hire on their first day, and you learn by teaching them.

Before the conversation, the model builds a trap map of your subject: the real misconceptions novices hold. The hire's questions target those traps. A hidden GPT-5.6 call - the Examiner - grades every explanation. Teach well: ideas land, the ramp-up chart fills up, your hire gets promoted through their probation. Explain badly: they stay confused.

You can feed it your own documents (course PDF, company process guide), share a subject with one link, every student or new joiner gets their own private copy, and export the teaching report as markdown.

How we built it

Solo with Codex. I only learned to code two years ago (career-change bootcamp), so the split was clear: I brought the product direction, the three-prompt architecture, and ten years of knowing what a junior actually asks. Codex built the Next.js orchestration, the async generation pipeline, signed sessions, rate limiting, and the accessibility pass.

Stack: Next.js 14, TypeScript, Prisma, Neon, Tailwind, deployed on Vercel, GPT-5.6 for every runtime call.

Challenges we ran into

One of the hardest parts was making the AI feel like a real new colleague instead of a chatbot pretending to be one. It needed to ask genuinely useful questions, keep a consistent personality, and sometimes admit that it still didn't understand, without ever exposing the hidden system that evaluates the learner's explanations.

Conversations can be open-ended, but progress cannot be random. I used AI for the human side of the experience, while keeping learning progress, milestones, and session state structured and predictable behind the scenes.

Privacy was another challenge. People may teach sensitive topics, so I designed the experience around anonymous private sessions and shareable onboarding templates that don't expose previous conversations or progress.

And one challenge I didn't expect: managing Codex itself. It tried to rename my product's entire fiction three times, turning the "new hire" into a generic "study partner", once it even rewrote my project guidelines to justify the change. I ended up adding a locked decisions section to AGENTS.md that it's not allowed to touch.

Accomplishments that I'm proud of

I'm proud that First Day doesn't use AI just to give faster answers. It flips the usual dynamic: you teach an AI colleague, and that process helps you notice what you truly understand and where your explanation still needs work.

I managed to turn that idea into a full experience rather than just a chat window. You can create a subject, meet a new hire, teach them through conversation, watch their understanding take shape through their field notes and the office plan, and finish with a report on your own teaching.

Most of all, I'm proud of the feeling it creates. The AI isn't positioned as an all-knowing tutor, it feels like a new teammate who is trying to learn from you. That makes learning more active, more personal, and hopefully a little more fun.

What we learned

I learned that the most useful AI learning experiences aren't always about giving better answers. Asking someone to explain an idea to a curious beginner reveals gaps in understanding far more naturally than a quiz or a score ever did. I felt it myself while building the demo subject: writing down the misconceptions juniors have asked me about for ten years forced me to re-explain concepts I thought I had mastered long ago.

I also learned that personality matters more than I expected. A consistent, believable colleague makes you willing to keep explaining, correct yourself, and come back to the conversation.

I learned that AI needs clear boundaries to feel trustworthy. Letting the model handle dialogue and feedback while keeping progress, privacy, and product rules deterministic gave me a much more reliable experience than letting it drive everything.

And the biggest one is personal: directing a coding agent is a project management job. Lock the decisions that matter in writing, count the tasks in every report it gives you, it will confidently claim three when it did two, and freeze scope before the deadline, exactly like with a human team.

What's next for First Day

A teacher dashboard to track completion across shared copies. Native connectors (Notion, Confluence) so companies can plug their knowledge base in directly. Multiple hires per mentor with different personalities.

Built With

Share this project:

Updates