diff --git a/cmd/claude-setup/main.go b/cmd/claude-setup/main.go
index 06e9b33..f774f3b 100644
--- a/cmd/claude-setup/main.go
+++ b/cmd/claude-setup/main.go
@@ -13,6 +13,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime/debug"
 	"strings"
 )
 
@@ -25,8 +26,21 @@ var apiContent string
 //go:embed skill/pitfalls.md
 var pitfallsContent string
 
-const version = "3"
-const versionMarker = "<!-- samurai-skill-v"
+// skillVersion returns the samurai module release this binary was built from
+// (the git tag, e.g. "v0.4.0"), used to mark and detect installed skill files.
+// When run from a local checkout (no module version), it falls back to "dev".
+func skillVersion() string {
+	if info, ok := debug.ReadBuildInfo(); ok && info.Main.Version != "" && info.Main.Version != "(devel)" {
+		return info.Main.Version
+	}
+	return "dev"
+}
+
+// skillMarker formats the HTML comment stamped into installed SKILL.md files,
+// used to detect whether an installed skill matches this binary's release.
+func skillMarker(version string) string {
+	return fmt.Sprintf("<!-- samurai-skill-%s -->", version)
+}
 
 func main() {
 	dir, err := os.Getwd()
@@ -41,10 +55,12 @@ func main() {
 	skillDir := filepath.Join(dir, ".claude", "skills", "samurai")
 	skillFile := filepath.Join(skillDir, "SKILL.md")
 
+	version := skillVersion()
+
 	// Check if already installed and up to date.
 	if existing, err := os.ReadFile(skillFile); err == nil {
-		if strings.Contains(string(existing), versionMarker+version+" -->") {
-			fmt.Println("samurai skill is already up to date (v" + version + ")")
+		if strings.Contains(string(existing), skillMarker(version)) {
+			fmt.Printf("samurai skill is already up to date (%s)\n", version)
 			return
 		}
 		fmt.Println("Updating samurai skill...")
@@ -56,7 +72,7 @@ func main() {
 
 	// Write all skill files.
 	files := map[string]string{
-		"SKILL.md":    skillContent + "\n" + versionMarker + version + " -->\n",
+		"SKILL.md":    fmt.Sprintf("%s\n%s\n", skillContent, skillMarker(version)),
 		"api.md":      apiContent,
 		"pitfalls.md": pitfallsContent,
 	}
diff --git a/cmd/claude-setup/skill/SKILL.md b/cmd/claude-setup/skill/SKILL.md
index 3d8d4db..f3c6143 100644
--- a/cmd/claude-setup/skill/SKILL.md
+++ b/cmd/claude-setup/skill/SKILL.md
@@ -5,6 +5,9 @@ description: "Samurai scoped testing framework for Go (github.com/zerosixty/samu
 
 # Samurai Test Writing Rules
 
+**Good For:** Go tests with a branching state-mutating tree (≥3 leaf paths, mutually exclusive siblings) using `samurai.Run` / `samurai.RunWith`.
+**Bad For:** read-only chains, sequential accumulation, same-action-different-input variants, flat independent tests (see "When to Use" below for full criteria).
+
 ## Rules
 
 - One action + its assertions = one `Test()`. Don't split assertions into child Tests — assert the result in the same callback that produced it. Child Tests are for new actions (setup, mutations, queries), not for checking fields of a parent's result
@@ -54,9 +57,7 @@ Common false positives: `_HappyPath` / `_AccessDenied` siblings (access-denied u
 
 Cost: samurai re-executes the full path per leaf — same cost as manually duplicating setup, just automated. Parallel execution offsets wall-clock time.
 
-## Additional resources
-
-- For the full API example (samurai.Run) and RunWith (custom context), see [api.md](api.md)
-- For validation rules (panic conditions) and wrong patterns, see [pitfalls.md](pitfalls.md)
+## Additional resources (lazy-load — read only when triggered)
 
-<!-- samurai-skill-v3 -->
+- **Read [api.md](api.md) only when** you need the worked-out `samurai.Run` / `RunWith` API example (e.g. wiring a new test scope, writing a custom context type for `RunWith`). The "When to Use" / "Detection protocol" sections above are sufficient for review and detection work.
+- **Read [pitfalls.md](pitfalls.md) only when** debugging a panic, a validation error, or a wrong pattern (e.g. asserting on a parent's result inside a child Test, declaring vars where they should be assigned). Skip otherwise — the rules above already cover the common authoring decisions.