Direction Before Destination

Think about how a relationship between two people begins. Nobody sits down before a first conversation and writes out what a friendship's roadmap should contain or which other people it should lead to. What exists at the start is simply a sense of curiosity.

I want to take that ordinary experience seriously, because I think it describes how a great deal of meaningful work proceeds, including research and product engineering.

We start not knowing what to build, sometimes without even a direction. If we just take a first step rather than standing still planning, things begin to happen. A small thing we build shows us a second worth building. The direction narrows until we have a pathway; we continue along it and eventually we arrive at a milestone, although we could not have named it at the start.

This is the fog of war: in a strategy game the map is dark where none of your units has been, and movement is the instrument of seeing. The metaphor understates it, though. In a game the terrain exists before you reach it. In a relationship, some of what you come to know did not exist at all until the going created it.

I think research and development can have the same structure. For example, let's look at NotebookLM.

Raiza Martin, its product lead, has described how it began as a 20% project (Google's program for self-directed work). One engineer had been working on something called "talk to small corpus," using a language model to converse with your own data. Talking with him sent her back to her experience as an adult learner and to the idea of talking to a textbook. Project Tailwind was that idea made concrete: you uploaded a PDF and chatted with it. Google's own retrospective states the wider aim, helping people research, think and write from sources they already had on hand. The first prototype took six weeks, with four or five people working part time. Early users, writing in through in-app forms and a Discord server, prompted additions: saving responses as notes, suggested questions, inline citations.

The team had already been testing audio. They made a monologue and a dialogue, and the dialogue was what people responded to, because it sounded like a podcast rather than text read aloud. That became Audio Overviews, two voices discussing the sources you supply, and Martin says its reception surprised them.

Quite often I'm surprised myself by how things unfold for me. Hence, I find myself engaged in at least one 20% project alongside my main work. It gives me fresh ideas and allows me to look beyond what immediacy offers.

We set direction early, but we add details as we move along. What do we do when we don't even know which direction to set? I have a playbook for that as well:

1. Find a source of inspiration, such as a post on LinkedIn discussing a novel approach.
2. Remember a project you worked on that you vaguely associate with the newfound inspiration.
3. Pull both of these sources (e.g., repositories of your project and the third-party tool) into the same workspace.
4. Narrate to Antigravity one thing you care about in the context of your project.
5. Send it to one or multiple agents in /plan mode.
   …
6. Come back, read the reports and remove the noise, keeping only one interesting finding.
7. Build something with or write about the finding.
8. Go back to any of the previous steps and repeat to get new insights.

Coding harnesses like Claude Code, Codex and Google Antigravity are exceptionally good at divergent analysis, but the trick is to engineer the context in a way that makes seeing new insights more likely.