Dumping the Codebase for NotebookLM

Dumping the codebase to NotebookLM as a substitute for external review when there's no team. LLMs are warped in known ways; knowing how the surface bends is what makes the reflection useful.

How do you know if your work reads the way you think it reads?

Working alone means there's no second pair of eyes by default. Code review in a team surfaces this naturally — someone reads without having written it, and what surprises them is information. Solo, that loop doesn't close on its own.

NotebookLM is a reasonable substitute: upload your sources, ask questions, generate a podcast that synthesizes the material. But calling it an external read misses something. It's not a flat mirror. It's a fun house mirror: one where the surface is warped in known directions.

LLMs are warped in known ways. They latch onto what's legible and repeat it back. They skip what didn't land without telling you it missed. They flatten things you thought were sharp. None of this is a surprise. Working with them long enough, you develop a feel for which way the surface bends. That's what makes it useful rather than just noisy.

The warp is the instrument. If it takes something minor and over-indexes on it, two readings: you've said it more than you realized, so repetition made it prominent. Or the work around it isn't doing enough for that thing to actually register as minor. Neither feels great, but both are information. The proportionality problem was already there. The warp just made it visible.

If you work with coding agents, there's a second use. What keeps surfacing in the NotebookLM output is close to what an agent will reach for first when it needs context. A concept that comes up repeatedly in the podcast will come up repeatedly in your sessions. Seeing it early lets you decide whether that's the attachment you want, or whether the emphasis in the source material needs adjusting before you hand it to an agent.

For a portfolio the sources split naturally: architecture documents, source code, published content. A script in this project assembles them into flat files ready for upload. Run it after a sprint, after a structural change, after publishing something new — not to check if it's good, but to see what the warp surfaces.

The content is already prose. The tool collects it without transformation. The interesting part isn't the tool.