diff --git a/.claude/notes/automation.md b/.claude/notes/automation.md
index 0bb32b54..fbd5c3f9 100644
--- a/.claude/notes/automation.md
+++ b/.claude/notes/automation.md
@@ -3,6 +3,35 @@
 Lessons specific to Claude Code skills, slash commands, classifier
 interactions, and parallel-agent workflows in this repo. Newest-first.
 
+### Match shell-command hooks by regex word-boundary, not substring
+
+Substring patterns like `*"gh pr create"*` fire on any command that
+literally contains the target string (`echo gh pr create`,
+`bash -c "git push test"`), producing high-volume false positives.
+"Starts with" patterns miss compound forms (`cat > body && gh pr
+edit`, `var=$(...) && git push`), producing high-cost false
+negatives. A regex word-boundary match lands the balance — three
+parts: a leading shell-separator boundary, the alternation, a
+trailing boundary:
+
+```
+(^|[[:space:];&|])     # line start OR shell separator
+(gh pr create|git push) # the actual targets
+([[:space:]]|$)        # space OR line end
+```
+
+For advisory hooks (surface a checklist, log a warning) bias toward
+false positives — the cost of an extra fire is one acknowledgment;
+a missed fire on a real action is worse. For blocking hooks (deny
+the action via `permissionDecision: deny`) flip the bias: prefer
+starts-with so an unrelated command isn't forcibly stopped.
+
+This trade-off generalizes to any agent hook matching shell commands
+by content; PR commands are this entry's example, not its scope.
+
+Anchor: `.claude/settings.json` PR #80 commit `3c12d02`,
+`PreToolUse.Bash.hooks[0].command` regex.
+
 ### Don't escape backticks inside a single-quoted HEREDOC
 
 A `bash -c '...$(cat <<'EOF' ... EOF)...'`-style PR-body payload uses
diff --git a/.claude/notes/skill-design.md b/.claude/notes/skill-design.md
index 92c2cca5..bf49ad78 100644
--- a/.claude/notes/skill-design.md
+++ b/.claude/notes/skill-design.md
@@ -3,6 +3,27 @@
 Lessons specific to authoring `.claude/skills/<name>/SKILL.md` files
 and related workflow scaffolding. Newest-first.
 
+### Orchestrate multi-phase skills directly, not through plugin $ARGUMENTS
+
+Routing a skill's full prompt body through a plugin command's
+$ARGUMENTS blows the Claude Code permission classifier when the
+prompt grows past ~17 KB ("parser timeout, resource limit, or
+over-length"). The original `pr-review-loop` skill delegated its
+5-phase workload to a plugin's setup script and hit this wall in
+real use; PR #73 dropped the dependency and re-routed the phases
+to direct Agent-tool dispatches.
+
+For any new multi-phase skill with a large prompt body, orchestrate
+inline. Plugin delegation remains fine when `$ARGUMENTS` stays well
+under ~1 KB and the protocol is well-typed (small command + flags);
+this lesson is about the multi-phase-with-large-prompt shape that
+hits the classifier wall, not plugins in general. If a plugin is
+the right shape but the prompt is heavy, pass it by file path —
+never the whole body as one positional arg.
+
+Anchor: PR #73 commit `6ed4fa8`, `.claude/skills/pr-review-loop/SKILL.md`
+"How the loop runs" section.
+
 ### Mechanism-test a skill before declaring it ready
 
 Manual walkthrough of a multi-step skill protocol misses substitution