AI-Assisted Product Workflow
Field notes on using AI tools through the product process. Distilled from working as a PM at an early-stage startup.
A working pipeline: context → prototype → build
- Notebook-style LLM for context — load the background, research, and prior docs into an LLM you can interrogate, so you’re reasoning over real context instead of memory.
- AI design tool for prototypes — generate quick prototypes/mocks to make the idea tangible and get reactions fast.
- Build the real thing in a design tool (e.g. Figma) — once the prototype has shape, produce the actual design artifact engineering builds from.
Each step de-risks the next: context sharpens the prototype, the prototype sharpens the real design.
Vibe-coding works once you can write the whole flow
“Vibe coding” with an AI assistant is genuinely good — but only once you can articulate the entire flow end to end. The AI fills in implementation; you must own the logic. If you can’t write the flow out, you’re not ready to hand it to the model. (This is the same skill that makes a good PRD — see Writing-PRDs.)
Keep judgement in an AI world
AI makes producing output cheap, which makes judgement the scarce skill. The value isn’t generating ten options — it’s knowing which one is right, what to cut, and when “good enough” is actually good enough.
Learn by playing
Play with codebases and an AI assistant directly. Poking at real code with a model that explains it is one of the fastest ways to build technical intuition (see Engineering-Practices).