diff --git a/apps/web/src/services/ai/orchestrator-sub-agent.ts b/apps/web/src/services/ai/orchestrator-sub-agent.ts
index 35b5a0079..db0656ef7 100644
--- a/apps/web/src/services/ai/orchestrator-sub-agent.ts
+++ b/apps/web/src/services/ai/orchestrator-sub-agent.ts
@@ -451,6 +451,23 @@ async function executeSubAgent(
     reducedComplexity,
   );
 
+  // When element tools are enabled, the ELEMENT_TOOL_OUTPUT_FORMAT block
+  // (appended below) is the authoritative output-format instruction —
+  // teaching the model to emit `<op_tool>` tags. The jsonl-format /
+  // jsonl-format-simplified skills lead with a CRITICAL directive saying
+  // "Output ONLY ```json … Do NOT use tool calls", which directly conflicts
+  // with `<op_tool>` and confuses weak models (MiniMax / GLM / Kimi):
+  // they anchor on the front-of-prompt CRITICAL and emit raw JSONL,
+  // bypassing every element tool — which defeats the entire point of the
+  // n-tools-per-element design (stability for weak models in the built-in
+  // sub-agent path). Drop those two skills here so ELEMENT_TOOL_OUTPUT_FORMAT
+  // is the only output-format instruction the model sees.
+  if (elementToolsEnabled) {
+    resolvedSkills = resolvedSkills.filter(
+      (s) => s.meta.name !== 'jsonl-format' && s.meta.name !== 'jsonl-format-simplified',
+    );
+  }
+
   const skillPrompt = resolvedSkills.map((s) => s.content).join('\n\n');
   const systemPrompt = elementToolsEnabled
     ? skillPrompt + '\n\n' + ELEMENT_TOOL_OUTPUT_FORMAT
diff --git a/packages/pen-ai-skills/skills/phases/generation/elements.md b/packages/pen-ai-skills/skills/phases/generation/elements.md
index e2d143437..5be1229f0 100644
--- a/packages/pen-ai-skills/skills/phases/generation/elements.md
+++ b/packages/pen-ai-skills/skills/phases/generation/elements.md
@@ -10,14 +10,16 @@ category: base
 ---
 
 <!--
-  IMPORTANT: This skill is gated by the `hasMcpTools` flag. It only
-  auto-loads into the generation-phase prompt when the caller declares
-  the AI has live access to MCP element tools (external clients:
-  Claude Code / Codex / Gemini CLI / Cursor). The embedded orchestrator
-  in apps/web emits single-shot JSON and cannot call MCP tools — this
-  skill would be 1500 tokens of dead weight there, so it stays excluded.
+  IMPORTANT: This skill is gated by the `hasMcpTools` flag. It auto-loads
+  into the generation-phase prompt for any caller that declares element-
+  tool access — both external MCP clients (Claude Code / Codex / Gemini
+  CLI / Cursor) AND the apps/web embedded orchestrator when
+  VITE_ENABLE_ELEMENT_TOOLS is on. In the embedded path the model emits
+  `<op_tool>` blocks (parsed by `tryParseAllElementToolOutputs` and
+  dispatched via `element-tools-dispatcher.ts`), giving weak models the
+  same schema-locked surface as external MCP clients.
 
-  External MCP clients still retrieve the content explicitly via
+  External MCP clients can also retrieve this content explicitly via
   get_design_prompt(section='elements'), which bypasses resolveSkills'
   trigger filter (uses getSkillByName for direct lookup).
 
@@ -50,8 +52,6 @@ Multi-tool example — a "Notifications" settings section with a header + 4 togg
 
 ## Theme handling — when to pass `theme: 'system'`
 
-> **This section applies to the MCP tool-call path only** (codex CLI / Claude Code / Cursor calling `add_X_v1` directly). The web-app sub-agent JSONL path forbids tool calls — there, write `$color-*` / `$type-*` / `$spacing-*` / `$radius-*` refs directly in JSONL fills/fontSize/etc. See the **DESIGN SYSTEM TOKENS** section in `jsonl-format.md` for the available token names and when refs vs literals are appropriate.
-
 **Default to `theme: 'system'`** for every v1 tool call. This makes the output respect the user's design system (`doc.variables` / `doc.themes`):
 
 - If user has seeded `applySemanticPalette(doc)` or set custom token values, v1 'system' mode emits `$color-*` / `$type-*` / `$spacing-*` / `$radius-*` refs that resolve to user's design system at paint time.