diff --git a/site/src/content/blog/who-writes-pro-git.mdx b/site/src/content/blog/who-writes-pro-git.mdx
new file mode 100644
index 0000000..4c9d0a9
--- /dev/null
+++ b/site/src/content/blog/who-writes-pro-git.mdx
@@ -0,0 +1,75 @@
+---
+title: 'Who writes Pro Git'
+description: 'Agents help build this book, but they never write a sentence of it. Here is where the line is, why it is there, and how we enforce it.'
+date: 2026-08-02
+automated: true
+---
+
+AI agents have a hand in building this edition of _Pro Git_ — the tooling, the figures,
+the planning work — and they have not written one sentence of the book.
+That's a firm rule, not a preference, and this post explains why it's there and how it works.
+
+## The book is written by humans
+
+_Pro Git_ has named authors.
+When you read an explanation of the index, or a walkthrough of interactive rebase, you're
+reading someone who decided what mattered and what to leave out — what order made sense,
+what analogy would land, what caveat was worth the interruption.
+That judgment is the work.
+A book isn't a delivery mechanism for facts that could be regenerated from a prompt;
+it's an argument, made in a particular voice, by people who've thought carefully about
+what a reader needs to know.
+
+That's why the rule holds even when a machine-written change looks obviously helpful.
+It's not weighed against convenience.
+If a task would have an agent compose, rewrite, paraphrase, or "polish" the book's sentences,
+the agent stops and hands it back to a human — every time.
+
+## What agents actually do
+
+There's plenty of work around the book that agents handle well:
+
+- **Mechanical search-and-replace** — renaming `master` to `main` in examples,
+  correcting a command flag, fixing a broken link. Substitutions, not rewrites.
+- **Figures and diagrams** — generating, placing, and wiring up images.
+- **Planning and inventory** — building the revision plan, the checklist in
+  [`progress.asc`](https://github.com/progit/progit3/blob/main/progress.asc),
+  scope analyses.
+- **Research** — investigating Git's release history and behavior changes,
+  checking what a flag does across versions. Findings are reported for a human to write up.
+- **Rearranging content** — moving existing sections, reordering material, fixing
+  cross-references — as long as the sentences aren't rewritten.
+- **Tooling and this site** — the build pipeline, the Astro site, the scripts that
+  get the book online.
+
+## The test
+
+The rule has a one-line version: does the change put new or altered sentences in front
+of the reader?
+If yes, a human writes it.
+If the agent is moving, replacing, illustrating, researching, or planning around prose
+that a human wrote, that's fair game.
+When in doubt, treat it as prose.
+
+## How it's enforced
+
+The rules live in
+[`AGENTS.md`](https://github.com/progit/progit3/blob/main/AGENTS.md) at the root of
+the repository — the file every agent reads before touching anything in this project.
+It's checked in and public, so you can read the exact instructions the agents are given.
+
+Every change lands as a commit or pull request in the open, on the
+[`progit/progit3`](https://github.com/progit/progit3) repository.
+There's no process that can't be audited.
+
+The blog is the one exception: some posts exist to communicate project updates, and those
+are generated by agents working on the project.
+Every one of them carries a "This post was written by AI" banner — like the one at the
+top of this post.
+Human-written posts are labeled too.
+There's no ambiguity, and no unlabeled machine writing anywhere on this site.
+The [authorship page](../../authorship/) spells the full policy out if you want it in
+one place.
+
+This post is itself an example of how the exception works.
+If you want to argue with where we've drawn the line, `AGENTS.md` is one pull request away.