diff --git a/data/skills/image_translation.md b/data/skills/image_translation.md
index 90201ad..72f7a99 100644
--- a/data/skills/image_translation.md
+++ b/data/skills/image_translation.md
@@ -38,7 +38,9 @@ attempting to recover artwork from a flattened text-bearing PNG.
 
 Translate embedded bitmap text by reconstructing the smallest safe UI regions and rendering
 approved target text. Treat each bitmap as a structured interface asset, not merely an OCR
-surface. Do not use generative image editing or inpainting.
+surface. Do not use generative editing or inpainting on a complete asset. A narrow exception for
+generating an isolated stylized title wordmark is defined below; all background reconstruction,
+placement, and final compositing must remain deterministic.
 
 ## Required outcome
 
@@ -249,6 +251,38 @@ asset. Preserve character-name colors and other meaningful visual distinctions.
 font is unavailable, choose the closest font by shape and weight, but still match the original
 effect stack and visual prominence.
 
+### 10a. Handle stylized title wordmarks
+
+Use ordinary deterministic typography by default. When a title or logo depends on hand-shaped,
+dimensional lettering that available fonts cannot reproduce faithfully, image generation may be
+used only for the isolated replacement wordmark and only when the user has authorized generative
+output.
+
+1. Load the original title image as a style reference, not as a generative edit target.
+2. Request one isolated wordmark containing the exact translated text on a perfectly flat,
+   removable chroma-key background. Require generous padding and prohibit characters, scenery,
+   taglines, badges, watermarks, extra words, and every other source-image element.
+3. State the target text verbatim and reject the output if any letter, punctuation mark, or word
+   differs. Do not repair a misspelling by hiding or painting over part of the generated logo.
+4. Match the source logo's letter-face colors, gradient, highlights, outline stack, dimensional
+   shadow, energy, and overall silhouette. Do not ask the generator to reproduce covered artwork.
+5. Remove the flat key to alpha with a deterministic chroma-key tool. Validate transparent
+   corners, plausible subject coverage, crisp antialiased edges, and absence of key-color fringe.
+6. Crop to the alpha bounds, scale with a high-quality deterministic filter, and position it from
+   the source logo's measured bounds and runtime layout. Keep every protected artwork region and
+   separate label outside the generated step.
+7. Remove the source wordmark independently using a clean source layer or another already-approved
+   deterministic reconstruction. Never rely on the new logo to conceal old glyphs; inspect for
+   source-text fragments above, below, and between the replacement letters at original resolution.
+8. Composite the extracted wordmark deterministically onto a candidate rendered from the verified
+   original or backup. Reuse the same accepted wordmark for every title variant instead of
+   regenerating inconsistent copies. Translate taglines, demo labels, and other simple text
+   deterministically.
+9. Compare the candidate with the source and a deterministic typography candidate at original and
+   runtime scale. Reject it if it alters the background, obscures meaningful art, introduces a
+   backing panel, looks less faithful, or leaves source-glyph ghosts. Do not install a generated
+   candidate until the user has visually approved it.
+
 ### 11. Protect dynamic values
 
 Reserve the exact runtime-drawn areas discovered from available project or runtime data. Test likely