diff --git a/.pi/skills/grill b/.pi/skills/grill
new file mode 120000
index 0000000..2806aae
--- /dev/null
+++ b/.pi/skills/grill
@@ -0,0 +1 @@
+../../src/skills/grill
\ No newline at end of file
diff --git a/AGENTS.md b/AGENTS.md
index 169ea08..bb3f712 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -19,6 +19,8 @@
 
 ## Pull requests
 
+- Always create a new branch before opening a pull request.
+  Never open a pull request directly from a shared or existing work branch.
 - During pull-request development, add follow-up commits instead of amending existing commits.
   Preserve review history until the pull request is ready to merge.
 - When the user says a pull request is merged, clean up the local branch without waiting for an explicit request.
diff --git a/src/skills/grill/SKILL.md b/src/skills/grill/SKILL.md
new file mode 100644
index 0000000..2d4b583
--- /dev/null
+++ b/src/skills/grill/SKILL.md
@@ -0,0 +1,82 @@
+---
+name: grill
+description: Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any grill trigger phrase.
+---
+
+# grill
+
+Interview the user relentlessly until you reach a shared understanding.
+Map the interview as a **design tree**.
+Every decision branches into the decisions that depend on it.
+
+Never use an interactive question tool.
+Do not act on the plan until the user confirms shared understanding.
+Do not begin implementation after confirmation unless the user explicitly asks for implementation.
+
+## Steps
+
+1. Build the design tree.
+   Identify the root decision and every known dependent decision.
+   Read relevant files, docs, and prior conversation before asking.
+   Resolve facts yourself when the environment can answer them.
+
+   Finish when known prerequisites, facts, and open decisions are mapped.
+
+2. Freeze the current frontier.
+   Treat the **frontier** as every decision whose prerequisites are settled.
+   Freeze those questions into the current round.
+   Put newly unblocked questions in the next round.
+   Put dependent questions in a later round when their prerequisite remains unsettled.
+
+   Finish when the current round contains every currently unblocked decision.
+
+3. Ask one question from the current round.
+   Ask only one question per assistant turn.
+   Number the question.
+   Include your recommended answer.
+
+   When a question turns on existing code, docs, or artifacts, open with a context quote.
+   Quote the exact passage that creates the question.
+
+   Finish when the user can answer without scrolling through other questions.
+
+4. Wait for the user's answer.
+   Treat decisions as the user's job.
+   Treat facts as your job.
+   Do not ask the user for facts that tools or files can answer.
+
+   Finish when the answer updates the design tree.
+
+5. Continue or advance the round.
+   Ask the next unanswered question from the frozen round.
+   Recompute the frontier only after the frozen round is complete.
+   Challenge vague or overloaded language with precise terms.
+   Probe edge cases with concrete scenarios.
+   Surface contradictions between user claims, docs, and code with quotes from both sides.
+
+   Finish when the next turn has either another question from the current round or a new frozen round.
+
+6. Repeat rounds until the frontier is empty.
+   Stop only when every branch of the design tree has been visited.
+   Leave no decision silently assumed.
+
+   Finish when you have visited every branch and no open decision remains.
+
+7. Close the interview.
+   State that the frontier is empty.
+   Ask the user to confirm that you reached a shared understanding.
+   Stop and wait for the user's next instruction.
+
+   Finish when the user can confirm the shared understanding or give a new instruction.
+
+## Question format
+
+Use this format for each question:
+
+```markdown
+Q1. **<question title>**
+
+<question body, including concrete choices when helpful>
+
+Recommendation: <your recommended answer>
+```