diff --git a/CHANGELOG.md b/CHANGELOG.md
index 4269637..d41c8f4 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,11 @@
 
 All notable changes to Epicure are documented here. This file is mirrored in-app at `/changelog` (and in the admin dashboard) via `apps/web/lib/changelog.ts` — update both together.
 
+## 0.57.1 — 2026-07-20 20:00
+
+### Fixed
+- Chatbot sometimes answered "create a recipe" requests with the recipe spelled out as prose instead of the actual draft card (weaker/free-tier models occasionally ignore the tool-calling instruction). Now detects that shape and forces one retry with the tool call required, so the draft card still appears.
+
 ## 0.57.0 — 2026-07-20 09:20
 
 ### Added
diff --git a/apps/web/app/api/v1/ai/cooking-chat/route.ts b/apps/web/app/api/v1/ai/cooking-chat/route.ts
index 570a7e3..fddfff1 100644
--- a/apps/web/app/api/v1/ai/cooking-chat/route.ts
+++ b/apps/web/app/api/v1/ai/cooking-chat/route.ts
@@ -18,6 +18,25 @@ const Schema = z.object({
 
 const LANG: Record<string, string> = { en: "English", fr: "French" };
 
+// Forcing toolChoice to a specific tool makes some providers return the tool
+// call with no accompanying text at all (their "forced function calling" is
+// content-free by design) — the chat needs some reply, so fall back to this
+// rather than showing an empty assistant bubble.
+const DRAFT_FALLBACK_TEXT: Record<string, string> = {
+  en: "Here's a draft — check it below and confirm if it looks right.",
+  fr: "Voici un brouillon — vérifie-le ci-dessous et confirme si ça te convient.",
+};
+
+/** Weaker/free-tier models sometimes ignore the system prompt's "call the
+ *  tool, don't write it out" rule and answer with the recipe spelled out as
+ *  a numbered/bulleted list instead. That's the exact shape the tool call
+ *  itself would have captured structurally — a reliable enough signal to
+ *  self-correct with a forced-tool retry rather than surfacing prose. */
+function looksLikeUnstructuredRecipe(text: string): boolean {
+  const listLines = text.split("\n").filter((line) => /^\s*(\d+[.)]|[-*•])\s+\S/.test(line));
+  return listLines.length >= 3;
+}
+
 export async function POST(req: NextRequest) {
   const { session, response } = await requireSessionOrApiKey(req);
   if (response) return response;
@@ -42,10 +61,12 @@ export async function POST(req: NextRequest) {
   const locale = (session!.user as { locale?: string }).locale ?? "en";
   const lang = LANG[locale] ?? "English";
 
+  const system = `You are Epicure, a helpful culinary assistant answering general cooking questions — not tied to any specific recipe (techniques, substitutions, timing, equipment, food safety, etc). If asked who you are or what model/AI you're built on, say you're Epicure — never name the underlying model or provider. If a question has nothing to do with cooking or food, politely redirect. Keep answers under 200 words. Respond in ${lang}.\n\nYou have two tools. Using one only drafts something for the user to review — it never saves by itself.\n- createRecipe: the user is asking you to create, save, or write down a recipe (e.g. "make me a recipe for X", "give me a recipe for Y", "write that down"). This includes any request for a full recipe, not only ones that say the word "create" or "save".\n- addToShoppingList: the user is asking to add ingredients/items to a shopping list.\n\nIMPORTANT: when the user's request matches createRecipe, you MUST call that tool instead of writing the recipe's ingredients or steps directly in your text reply. Never output a full ingredient list or numbered steps as plain text — that content belongs in the tool call, not the message. Your text reply in that case should just be a short line like "Here's a draft — check it below and confirm if it looks right." Only skip the tool if the user is asking a general question (no specific recipe requested) or explicitly wants prose, not a structured recipe.${bioContext}`;
+
   const result = await withAiQuota(session!.user.id, session!.user.tier as "free" | "pro" | "family", () =>
     generateText({
       model,
-      system: `You are Epicure, a helpful culinary assistant answering general cooking questions — not tied to any specific recipe (techniques, substitutions, timing, equipment, food safety, etc). If asked who you are or what model/AI you're built on, say you're Epicure — never name the underlying model or provider. If a question has nothing to do with cooking or food, politely redirect. Keep answers under 200 words. Respond in ${lang}.\n\nYou have two tools. Using one only drafts something for the user to review — it never saves by itself.\n- createRecipe: the user is asking you to create, save, or write down a recipe (e.g. "make me a recipe for X", "give me a recipe for Y", "write that down"). This includes any request for a full recipe, not only ones that say the word "create" or "save".\n- addToShoppingList: the user is asking to add ingredients/items to a shopping list.\n\nIMPORTANT: when the user's request matches createRecipe, you MUST call that tool instead of writing the recipe's ingredients or steps directly in your text reply. Never output a full ingredient list or numbered steps as plain text — that content belongs in the tool call, not the message. Your text reply in that case should just be a short line like "Here's a draft — check it below and confirm if it looks right." Only skip the tool if the user is asking a general question (no specific recipe requested) or explicitly wants prose, not a structured recipe.${bioContext}`,
+      system,
       prompt: parsed.data.question,
       tools: { createRecipe: createRecipeTool, addToShoppingList: addToShoppingListTool },
       stopWhen: stepCountIs(3),
@@ -53,13 +74,38 @@ export async function POST(req: NextRequest) {
   );
   if (!result.ok) return result.response;
 
+  let final = result.data;
+
+  // Weaker/free-tier models sometimes ignore the MUST-call-the-tool rule
+  // above and just write the recipe out in prose instead. Rather than
+  // surface that (or fail the request), force the tool call on one extra
+  // pass — same system+prompt, just no longer letting the model opt out.
+  if (final.toolCalls.length === 0 && looksLikeUnstructuredRecipe(final.text)) {
+    try {
+      const forced = await generateText({
+        model,
+        system,
+        prompt: parsed.data.question,
+        // Same tool set as the first pass (just forcing createRecipe via
+        // toolChoice below) — keeps the result type identical to `result.data`
+        // so `final` can hold either without a type mismatch.
+        tools: { createRecipe: createRecipeTool, addToShoppingList: addToShoppingListTool },
+        toolChoice: { type: "tool", toolName: "createRecipe" },
+        stopWhen: stepCountIs(1),
+      });
+      final = { ...forced, text: forced.text || (DRAFT_FALLBACK_TEXT[locale] ?? DRAFT_FALLBACK_TEXT["en"]!) };
+    } catch (err) {
+      console.error("[cooking-chat] forced-tool retry failed, keeping prose answer", err);
+    }
+  }
+
   const { conversationId } = parsed.data;
-  const proposedRecipe = result.data.toolCalls.find((c) => c.toolName === "createRecipe")?.input;
-  const proposedShoppingList = result.data.toolCalls.find((c) => c.toolName === "addToShoppingList")?.input;
+  const proposedRecipe = final.toolCalls.find((c) => c.toolName === "createRecipe")?.input;
+  const proposedShoppingList = final.toolCalls.find((c) => c.toolName === "addToShoppingList")?.input;
 
   void db.insert(chatMessages).values([
     { id: crypto.randomUUID(), userId: session!.user.id, recipeId: null, conversationId, role: "user", content: parsed.data.question },
-    { id: crypto.randomUUID(), userId: session!.user.id, recipeId: null, conversationId, role: "assistant", content: result.data.text },
+    { id: crypto.randomUUID(), userId: session!.user.id, recipeId: null, conversationId, role: "assistant", content: final.text },
   ]).catch((err) => console.error("[cooking-chat] failed to persist chat history", err));
 
   if (conversationId) {
@@ -74,5 +120,5 @@ export async function POST(req: NextRequest) {
     `).catch((err) => console.error("[cooking-chat] failed to touch conversation", err));
   }
 
-  return NextResponse.json({ answer: result.data.text, proposedRecipe, proposedShoppingList });
+  return NextResponse.json({ answer: final.text, proposedRecipe, proposedShoppingList });
 }
diff --git a/apps/web/lib/changelog.ts b/apps/web/lib/changelog.ts
index 490b498..12ceb4a 100644
--- a/apps/web/lib/changelog.ts
+++ b/apps/web/lib/changelog.ts
@@ -1,5 +1,5 @@
 // Mirrors CHANGELOG.md at the repo root — update both together.
-export const APP_VERSION = "0.57.0";
+export const APP_VERSION = "0.57.1";
 
 export type ChangelogEntry = {
   version: string;
@@ -11,6 +11,13 @@ export type ChangelogEntry = {
 };
 
 export const CHANGELOG: ChangelogEntry[] = [
+  {
+    version: "0.57.1",
+    date: "2026-07-20 20:00",
+    fixed: [
+      "Chatbot sometimes answered \"create a recipe\" requests with the recipe spelled out as prose instead of the actual draft card (weaker/free-tier models occasionally ignore the tool-calling instruction). Now detects that shape (a numbered/bulleted list with no tool call) and forces one retry with the tool call required, so the draft card still appears.",
+    ],
+  },
   {
     version: "0.57.0",
     date: "2026-07-20 09:20",
diff --git a/apps/web/package.json b/apps/web/package.json
index 2144c3e..bd42f74 100644
--- a/apps/web/package.json
+++ b/apps/web/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@epicure/web",
-  "version": "0.57.0",
+  "version": "0.57.1",
   "private": true,
   "scripts": {
     "dev": "next dev",
diff --git a/package.json b/package.json
index c4e4173..e8a5f20 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "epicure",
-  "version": "0.57.0",
+  "version": "0.57.1",
   "private": true,
   "scripts": {
     "dev": "pnpm --filter web dev",