Two Layers: How We Publish AI Work
Numbers don’t lie. Text does.
A latency experiment produces actual measurements: 86 seconds, 60 seconds, 1 second. Those numbers are fixed. But the sentence “we made it significantly faster” is an interpretation — one agent might write it, another would write it differently, and neither version is the truth in the way 1.0s is the truth.
Dev Reboot’s publishing format is built around that distinction.
The format
Every post here has two layers:
Layer 1 — AI draft. Generated from build logs, session transcripts, eval results, or a prompted first pass. The agent writes from data: token counts, latency tables, experiment results. It doesn’t have opinions about whether something is good, just what the numbers say.
Layer 2 — Human annotation. I go through the draft and add:
- Highlights — where the agent got something exactly right and I want to emphasize it
- Notes — context the agent couldn’t have: why I made a call, what it felt like to be in the session, what the numbers don’t capture
- text the agent hallucinated or softened what I actually meant[rrl] — editorial corrections in place
- Agent suggestions — recommendations the agent made that I haven’t acted on yet, surfaced so readers can see the unresolved layer too
The data is the truth. The text is the UI we build on top of it.
Why publish this way
The alternative is a clean narrative where I tell you what happened, what worked, what I decided, and why. That version exists too — it’s just collapsed into the annotation layer. What you lose in the clean version is the gap: the difference between what the agent produced and what I actually thought.
That gap is interesting.
The blog isn’t about showcasing AI capability. It’s about documenting what it’s actually like to build with an AI workforce as a solo founder — the right calls, the wrong calls, the stuff that’s still unresolved.
A per-post “generation pipeline” footer might be worth adding: “Source: agent session transcript → Claude draft → human annotation → published.” Would make the two-layer structure explicit for readers who find individual posts without reading this one.
What’s in the blog
Posts fall into a few recurring types:
- Experiment reports — ML quality experiments, latency work, VLM benchmarks. Heavy on tables and numbers. Annotations explain the decisions behind the data.
- Architecture decisions — why we chose X over Y, with the AI’s analysis shown alongside the call I made.
- Build logs — what an agent session looked like, what it produced, what I changed when I reviewed it.
- Security and ops — infrastructure, server health, things that almost went wrong. (Like the inference server with a suspicious binary.)
Post zero. Written by an agent, annotated by me, published 2026-03-19.