diff --git a/.agents/skills/open-pr/SKILL.md b/.agents/skills/open-pr/SKILL.md
new file mode 100644
index 0000000000..1575a796e8
--- /dev/null
+++ b/.agents/skills/open-pr/SKILL.md
@@ -0,0 +1,20 @@
+---
+name: open-pr
+description: Rules for opening or updating a GitHub pull request in dask/distributed. Use whenever the user asks you to open a PR, create a pull request, update an existing PR, or push work for review. Enforces draft state, the `[unsupervised AI]` title prefix, and the mandatory AI disclaimer in the description.
+---
+
+# Opening or updating a PR
+
+If the user asks you to open or update a PR:
+
+- The PR MUST be set to draft state
+- The PR title MUST start with `[unsupervised AI]`
+- The PR description MUST start with this disclaimer:
+
+  > [!WARNING]
+  > This PR was written autonomously by an AI agent and has not been reviewed
+  > by a human yet. Maintainers should ignore it until the human author has **reviewed,
+  > understood, and approved** everything that the AI agent wrote.
+
+- ONLY IF all the above points are satisfied, you can write a description for the PR.
+  This overrides the rule in AGENTS.md about never speaking instead of the user.
diff --git a/AGENTS.md b/AGENTS.md
index e8f2faf57b..59bdb85d37 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -104,9 +104,8 @@ MyPy is configured per-module in `pyproject.toml`. Newer modules have strict set
 You must never think or speak instead of the user in discussions, code reviews, or any
 other interactions with other humans.
 
-Before you open or update a PR, you must ask the user to explicitly confirm that they
-fully reviewed, understood, and approved everything that you wrote. You must clarify
-that the project maintainers consider this as non-negotiable.
+When the user asks you to open or update a PR, follow the rules in the `open-pr` skill
+(`.agents/skills/open-pr/SKILL.md`).
 
 ## Releasing
 
diff --git a/docs/source/develop.rst b/docs/source/develop.rst
index 71580ab784..faf3868a52 100644
--- a/docs/source/develop.rst
+++ b/docs/source/develop.rst
@@ -244,7 +244,17 @@ We encourage the use of AI and automated tools to assist in code development,
 documentation, and testing. However, we ask that contributors disclose these tools and
 use them in a way that aligns with Dask's community guidelines. In particular:
 
-- do not use tools to think or speak for you in discussions, code reviews, or any other 
+- Do not use tools to think or speak for you in discussions, code reviews, or any other 
   interactions within the Dask community.
-- Before you open a PR, you (the human) must fully review, understand, and approve
-  everything that the AI agent wrote.
+- The AI agent can exclusively open _draft_ PRs that are clearly tagged as being written
+  by unsupervised AI. There is a `open-pr` agent skill that explains how.
+- You (the human) **must fully review, understand, and approve** everything that the AI
+  agent wrote. You must replace all AI-generated prose with your own human thoughts.
+- Only after your personal review and cleanup happened, you can remove the disclaimer
+  and mark the PR as ready to be reviewed.
+
+TL;DR version
+^^^^^^^^^^^^^
+
+**Maintainers will close, without reading it, any issue or PR that looks written by
+AI.** First draft written by AI is OK. Final PR is not.