diff --git a/packages/cli/src/commands/snapshot.ts b/packages/cli/src/commands/snapshot.ts
index 09c5a9362c..0c8972bf4a 100644
--- a/packages/cli/src/commands/snapshot.ts
+++ b/packages/cli/src/commands/snapshot.ts
@@ -139,17 +139,20 @@ async function captureSnapshots(
         })
         .catch(() => {});
 
-      // Wait for sub-compositions to be mounted by the runtime
-      // (they're fetched and injected asynchronously via data-composition-src)
+      // Wait for ALL sub-compositions to be mounted by the runtime.
+      // The old check resolved when the first sub-timeline registered, causing
+      // "last beat black" bugs: beat-5's sub-comp hadn't loaded yet when the
+      // snapshot seeked into its time range. Now we count data-composition-src
+      // host elements and wait until we have a matching number of sub-timelines.
       await page
         .waitForFunction(
           () => {
             const tls = (window as any).__timelines;
             if (!tls) return false;
-            const keys = Object.keys(tls);
-            // Wait until at least one sub-composition timeline is registered
-            // (not counting "main" or empty registrations)
-            return keys.length >= 2 || keys.some((k) => k !== "main");
+            const hosts = document.querySelectorAll("[data-composition-src]").length;
+            if (hosts === 0) return Object.keys(tls).length >= 1;
+            const subKeys = Object.keys(tls).filter((k) => k !== "main");
+            return subKeys.length >= hosts;
           },
           { timeout: timeoutMs },
         )
@@ -293,8 +296,19 @@ async function captureSnapshots(
             if (tls) {
               for (const key in tls) {
                 if (tls[key]?.seek) {
+                  // Sub-composition timelines run in local time relative to
+                  // their data-start. Seeking them to global time causes beats
+                  // with exit animations to appear black (global t clamps past
+                  // the exit). Compute local time: global_t - data_start.
+                  const host = document.querySelector<HTMLElement>(
+                    `[data-composition-id="${key}"]`,
+                  );
+                  const dataStart = host
+                    ? parseFloat(host.getAttribute("data-start") ?? "0") || 0
+                    : 0;
+                  const localTime = Math.max(0, t - dataStart);
                   tls[key].pause();
-                  tls[key].seek(t);
+                  tls[key].seek(localTime);
                 }
               }
             }
diff --git a/skills/website-to-hyperframes/references/step-4-vo.md b/skills/website-to-hyperframes/references/step-4-vo.md
index b0bd80eece..119edd8d3d 100644
--- a/skills/website-to-hyperframes/references/step-4-vo.md
+++ b/skills/website-to-hyperframes/references/step-4-vo.md
@@ -73,12 +73,18 @@ After the provider is selected, audition at least 2 voices with the first senten
     -H "x-api-key: $HEYGEN_API_KEY" | python3 -c \
     "import json,sys; v=json.load(sys.stdin)['data']; [print(x['voice_id'], x['name'], x['language']) for x in v[:10]]"
 
-  # Generate audio — returns audio_url + word_timestamps in one call
+  # Generate audio — response: { "data": { "audio_url": ..., "word_timestamps": [...] } }
   curl -s -X POST "https://api.heygen.com/v3/voices/speech" \
     -H "x-api-key: $HEYGEN_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{"text":"Your script here","voice_id":"VOICE_ID","speed":1.0}' \
-    | python3 -c "import json,sys; r=json.load(sys.stdin); print(r['audio_url']); open('transcript_raw.json','w').write(json.dumps(r.get('word_timestamps',[]),indent=2))"
+    | python3 -c "
+  import json,sys
+  r=json.load(sys.stdin)
+  d=r['data']
+  print(d['audio_url'])
+  open('transcript_raw.json','w').write(json.dumps(d.get('word_timestamps',[]),indent=2))
+  "
 
   # Then download the audio
   curl -sL "AUDIO_URL_FROM_ABOVE" --output narration.mp3
@@ -129,7 +135,16 @@ For ElevenLabs and HeyGen TTS, substitutions are usually unnecessary — they ha
 
 ## Transcribe for word-level timestamps
 
-**If you used HeyGen v3 TTS:** word timestamps were already returned in the generate call (`word_timestamps: [{ word, start, end }]`). Save them as `transcript.json` — the format matches what the beat mapping step expects. No separate transcription step needed.
+**If you used HeyGen v3 TTS:** word timestamps were returned in the generate call. Normalize the format before saving — HeyGen v3 uses `word` but the pipeline expects `text`:
+
+```python
+import json
+raw = json.load(open('transcript_raw.json'))
+normalized = [{"text": w["word"], "start": w["start"], "end": w["end"]} for w in raw]
+json.dump(normalized, open('transcript.json', 'w'), indent=2)
+```
+
+No separate transcription step needed.
 
 **If you used ElevenLabs or Kokoro:**