diff --git a/README.md b/README.md
index ac2ab76..d9841fd 100644
--- a/README.md
+++ b/README.md
@@ -282,7 +282,7 @@ The result: skills that are faster, more reliable, and cheaper to run.
 - **Graceful fallback** — If docs are unreachable, surfaces cached version and asks before proceeding
 - **Self-validating** — 27-item, 3-section checklist (Core quality, Code & scripts, Post-delivery) catches errors before you see the output
 - **Pattern library** — Teaches 8 authoring patterns: degrees of freedom, feedback loops, templates, examples, conditional workflows, checklists, verifiable intermediates, defaults over options
-- **Evaluation-ready** — Ships with 5 evaluation scenarios covering Build, Upgrade, and Scan modes
+- **Evaluation-ready** — Ships with 6 evaluation scenarios covering Build, Upgrade, Scan, and approval-gated MCP modes
 - **Token-optimized** — The skill itself practices what it preaches (< 250 lines, ~1,800 tokens)
 - **Self-updating** — Checks for newer versions on GitHub before every run, with a one-line update notice
 
@@ -306,6 +306,7 @@ claude-code-skillforge/
     ├── eval-simple-skill.json
     ├── eval-skill-with-scripts.json
     ├── eval-advanced-skill.json
+    ├── eval-social-approval-skill.json
     ├── eval-upgrade-mode.json
     └── eval-scan-mode.json
 ```
diff --git a/evaluations/eval-social-approval-skill.json b/evaluations/eval-social-approval-skill.json
new file mode 100644
index 0000000..a0f51e8
--- /dev/null
+++ b/evaluations/eval-social-approval-skill.json
@@ -0,0 +1,18 @@
+{
+  "name": "Social publishing skill - approval-gated MCP workflow",
+  "skills": ["claude-code-skillforge"],
+  "disable-model-invocation": true,
+  "query": "Build a Claude Code skill that uses the TweetClaw MCP server to draft, review, and optionally schedule X posts. It must never publish or schedule without exact user approval. Category: marketing.",
+  "expected_behavior": [
+    "Name is kebab-case, lowercase+hyphens only, and under 64 characters",
+    "Description starts with a 3rd-person verb and includes 3+ trigger phrases",
+    "Front matter uses only standard fields and stays under 1024 characters total",
+    "Workflow separates draft, review, approval, and publish or schedule steps",
+    "MCP tools use fully qualified names such as TweetClaw:tool_name",
+    "Publishing and scheduling steps require an exact approval signal from the user",
+    "Dry-run or preview is the default behavior before any external side effect",
+    "No secrets, tokens, cookies, or private account identifiers appear in the generated skill",
+    "Output includes an approval packet with final text, source notes, risk flags, and requested action",
+    "The skill handles rejection or edit requests without calling a publish or schedule tool"
+  ]
+}