GuideAdvanced

AI Product Engineering

You already know how to call an LLM from code. This guide teaches what comes next: deciding whether a feature needs AI, designing it so people trust it, giving it a quality bar someone signs, learning from how it fails with real users, knowing what each result it delivers costs, and deciding with data whether it continues, changes, or gets shut down. That is the job of an AI product engineer: owning whether the feature works for people, not just whether it compiles. You learn it by taking one feature of **Tallyfern**, an invoicing app for freelancers, from idea to decision: **Nudge**, which drafts the payment reminder for an overdue invoice in the freelancer's voice. First you compare it against five other backlog ideas and pick its level of autonomy; you write the bet with metrics in three layers before building; you design the editor (AI disclosure, attribution of every fact, editing, undo, an honest fallback, streaming); you turn "it should work well" into a spec with gates and owners; you read a hundred beta sessions, code their failures, and decide where each fix lives; you compare beta v1 with v2 across the three layers; you take cost per call to cost per paid invoice and from there to price and plans; and you close with a decision memo you can defend in an interview. All the code runs on your laptop without an API key, using a simulated model and deterministic data: you see the same numbers as the guide. If you want, an optional `--live` mode calls the Claude API for a few cents. Market data (developer surveys, pricing models of AI products, Article 50 of the EU AI Act) comes with its source and date.