diff --git a/tools/maintainer/review-pr-README.md b/tools/maintainer/review-pr-README.md
index f91a2acd..69efc73b 100644
--- a/tools/maintainer/review-pr-README.md
+++ b/tools/maintainer/review-pr-README.md
@@ -2,15 +2,53 @@
 
 Adversarial code review for GitHub PRs. Works with any LLM agent.
 
+> **Status: Experimental.** We're still figuring out how to use this effectively. Expect the workflow to evolve.
+
+## How It Works
+
+Point your agent at `review-pr.md` and ask it to review a specific PR:
+
+> "Read tools/maintainer/review-pr.md and apply it to PR #123"
+
+The tool will:
+
+1. Check out the PR branch locally
+2. Run an adversarial review (find at least 5 issues)
+3. Transform findings into professional tone
+4. Preview the review and ask before posting
+
+See `review-pr.md` for full prompt structure, severity ratings, and sandboxing rules.
+
+## When to Use
+
+**Good candidates:**
+
+- PRs with meaningful logic changes
+- Refactors touching multiple files
+- New features or architectural changes
+
+**Skip it for:**
+
+- Trivial PRs (typo fixes, version bumps, single-line changes)
+- PRs you've already reviewed manually
+
+## Workflow Tips
+
+**Always review before posting.** The preview step exists for a reason:
+
+- **[y] Yes** — Post as-is (only if you're confident)
+- **[e] Edit** — Modify findings before posting
+- **[s] Save only** — Write to file, don't post
+
+The save option is useful when you want to:
+
+- Hand-edit the review before posting
+- Use the findings as input for a second opinion ("Hey Claude, here's what Raven found — what do you think?")
+- Cherry-pick specific findings
+
+**Trust but verify.** LLM reviews can miss context or flag non-issues. Skim the findings before they hit the PR.
+
 ## Prerequisites
 
 - `gh` CLI installed and authenticated (`gh auth status`)
 - Any LLM agent capable of running bash commands
-
-## Usage
-
-Point your agent at `review-prompt.md` and ask it to review a specific PR:
-
-> "Read tools/maintainer/pr-review/review-prompt.md and apply it to PR #123"
-
-See `review-prompt.md` for full details on output format, safety features, and how it works.