Strong frame. The flip is that "dev tool" used to mean code editor — now it means anything that lets a non-engineer instruct an agent in natural language, hit run, and inspect output. Substack qualifies. Sheets qualifies. The TAM for "dev tools" just multiplied 100x without the category noticing.
The screen-record → HTML report flow nails the bottleneck most builders miss. It's not the input modality (typed vs voice vs video) — it's the reviewability of the agent's interpretation. Token cost is a red herring if the artifact catches one bad assumption before it compounds across three iterations.
The screen-record + voice loop solves the part of agent work text feedback can't reach: the tacit signal of 'this is what good looks like to me.' Been doing the same with Claude Code on my own builds and the compounding effect after 4-5 sessions is real. The agent starts inferring preferences from how you talk, not just what you say.
1/ long-time reader, first time caller. Thank you for all your writing!
2/ I started from your video --> html skill and built this out a little.
- I wanted to let the skill return a GIF if useful
- I wanted to have a quick review step
3/ It's working super well so far. I'm a total noob sharing and publishing skills, but here it is in case at all useful: https://github.com/jacobksamuelson/video-to-html
Strong frame. The flip is that "dev tool" used to mean code editor — now it means anything that lets a non-engineer instruct an agent in natural language, hit run, and inspect output. Substack qualifies. Sheets qualifies. The TAM for "dev tools" just multiplied 100x without the category noticing.
The screen-record → HTML report flow nails the bottleneck most builders miss. It's not the input modality (typed vs voice vs video) — it's the reviewability of the agent's interpretation. Token cost is a red herring if the artifact catches one bad assumption before it compounds across three iterations.
The screen-record + voice loop solves the part of agent work text feedback can't reach: the tacit signal of 'this is what good looks like to me.' Been doing the same with Claude Code on my own builds and the compounding effect after 4-5 sessions is real. The agent starts inferring preferences from how you talk, not just what you say.
Interesting...
I use a MacBook Pro, what video recording software do you recommend? Preferably free :)
Thanks!