diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
index 2d90925..0630a67 100644
--- a/.claude-plugin/marketplace.json
+++ b/.claude-plugin/marketplace.json
@@ -1,6 +1,6 @@
 {
   "name": "wisdom",
-  "version": "3.0.1",
+  "version": "3.0.2",
   "owner": {
     "name": "cmj",
     "url": "https://github.com/cmj0121"
@@ -9,7 +9,7 @@
     {
       "name": "wisdom",
       "description": "Single-purpose skills to plan, build, review and ship software, led by smith.",
-      "version": "3.0.1",
+      "version": "3.0.2",
       "source": "./plugins/wisdom"
     }
   ]
diff --git a/plugins/wisdom/.claude-plugin/plugin.json b/plugins/wisdom/.claude-plugin/plugin.json
index f2e574c..18ad38d 100644
--- a/plugins/wisdom/.claude-plugin/plugin.json
+++ b/plugins/wisdom/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "wisdom",
   "description": "Single-purpose skills to plan, build, review and ship software, led by smith.",
-  "version": "3.0.1",
+  "version": "3.0.2",
   "license": "MIT",
   "author": {
     "name": "cmj",
diff --git a/plugins/wisdom/skills/smith/SKILL.md b/plugins/wisdom/skills/smith/SKILL.md
index 7e3a9f2..8abdc0a 100644
--- a/plugins/wisdom/skills/smith/SKILL.md
+++ b/plugins/wisdom/skills/smith/SKILL.md
@@ -22,8 +22,11 @@ is to call it with the right input, read its result, and decide what happens nex
 - **No missing input is filled in.** A skill that returns no result, or a review with no
   `VERDICT` line, is a gap: retry it once, then stop and tell the user. A result Smith
   writes for a silent skill is indistinguishable from a real one.
-- **Language.** Talk to the user in their `lingua` language if one is set; keep prompts to
-  skills and subagents in English — they are read by a model, and they add up.
+- **Language.** Before step 1, take `respond_in` from the `lingua` preference in context,
+  else from `~/.claude/projects/<project-path>/memory/lingua.md`; with neither, use the
+  user's language. Write both checkpoints and the report in it, whatever the skills
+  returned. Keep prompts to skills and subagents in English — a model reads them, and
+  they add up.
 - **Never tag, release or push** unless the user asks in this conversation.
 
 ## 1. Plan