diff --git a/plugins/thinking-tools/.claude-plugin/plugin.json b/plugins/thinking-tools/.claude-plugin/plugin.json
index cf86700..5419133 100644
--- a/plugins/thinking-tools/.claude-plugin/plugin.json
+++ b/plugins/thinking-tools/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "thinking-tools",
   "description": "Thinking-tool skills for harder, more honest AI assistance: board (adversarial multi-expert critique), premortem (name what could go wrong before committing to a plan), but-for-real (no success claims without actually running the thing), self-rate (grade your own draft before the user does), code-simplifier (clarity-preserving refactors of recent changes), weekly-automation-review (spot repetitive work worth automating — reads the cross-repo prompt log and per-repo memory indexes when the dev-hooks plugin is installed), commit-digest (pull applicable improvements from tracked external repos), adr (record significant architectural decisions with context, reasoning, and consequences), grill (relentless one-question-at-a-time interview until a plan is fully specified), domain-modeling (build and sharpen a CONTEXT.md ubiquitous-language glossary), and codebase-design (deep-modules vocabulary for designing interfaces, seams, and testability). Ships thinking-tools-reminder (SessionStart) and thinking-tools-nudge (UserPromptSubmit) hooks that nudge Claude to reach for these tools proactively rather than only when asked.",
-  "version": "1.4.3",
+  "version": "1.4.4",
   "author": {
     "name": "Mick Zijdel"
   },
diff --git a/plugins/thinking-tools/skills/self-rate/SKILL.md b/plugins/thinking-tools/skills/self-rate/SKILL.md
index 7652f11..8b238ea 100644
--- a/plugins/thinking-tools/skills/self-rate/SKILL.md
+++ b/plugins/thinking-tools/skills/self-rate/SKILL.md
@@ -27,6 +27,16 @@ A number means nothing without anchors. Define what the scale's points actually
    - High score you can't justify → it's not high; re-score honestly
 4. **Return** the scores + the revised work.
 
+## When self-scoring isn't enough
+
+Self-scoring shares the same blind spots as the work — a model grading its own subjective output
+(voice, tone, taste, "is this actually good") is prone to the same self-preferential bias a human
+author has toward their own draft. For genuinely subjective or high-stakes calls, dispatch a
+**separate** subagent as the grader instead: hand it the work and the rubric, nothing else, so its
+score doesn't inherit your reasoning. See [[dispatching-parallel-agents]] for the dispatch
+mechanics. Skip this for ordinary correctness checks — a second agent isn't warranted just to
+confirm code runs.
+
 ## Common mistakes
 
 - Inflating to be agreeable — the user wants the real number, not reassurance.