diff --git a/skills/core/remotion.md b/skills/core/remotion.md
index 65be9d7..c0e7564 100644
--- a/skills/core/remotion.md
+++ b/skills/core/remotion.md
@@ -30,8 +30,10 @@ simple standalone operations that don't benefit from React rendering.
 | Video-only cuts with transitions | **Remotion** | Native `<OffthreadVideo>` + transitions |
 | Animated diagrams/text cards | **Remotion** | Frame-by-frame control |
 | Data-driven batch videos | **Remotion** | Zod props + parametric renders |
+| Word-level captions (in composition) | **Remotion** | CaptionOverlay with word highlight — superior to SRT |
+| Audio embedding (narration + music) | **Remotion** | Native `<Audio>` components with volume/fade |
 | Simple trim, concat (no composition) | FFmpeg | Instant, no Node dependency |
-| Subtitle burn-in (standalone) | FFmpeg | Proven, fast |
+| Subtitle burn-in (standalone, post-hoc) | FFmpeg | Only for adding subs to an already-rendered video without re-rendering |
 | Face enhance, color grade | FFmpeg | Filter-based, deterministic |
 | Remotion unavailable | FFmpeg | Automatic fallback |
 
@@ -328,13 +330,47 @@ Remotion renders are CPU-intensive but $0 API cost. Track via cost_tracker:
 - **Node.js 18+ required** — listed as optional in minimum system, required in recommended.
 - **Render in series, not parallel** — unless the machine has enough RAM. Each render spawns a Chromium instance.
 
+## Post-Render Verification Protocol (ALL pipelines)
+
+**Every Remotion render MUST be verified before presenting to the user.** This protocol applies
+to ALL pipelines, not just explainer. Pipeline-specific compose-directors may extend it but
+must not skip any step.
+
+**Step 1: Probe the output file (GATE — blocks all other steps):**
+```bash
+ffprobe -v quiet -print_format json -show_format -show_streams rendered_video.mp4
+```
+Verify ALL of:
+- [ ] Video stream exists with correct resolution and FPS
+- [ ] **Audio stream exists** — if missing, STOP immediately, fix audio config, re-render
+- [ ] Duration within ±5% of target
+- [ ] File size is reasonable (not 0 bytes, not suspiciously small)
+
+**If audio stream is missing, do NOT proceed.** This means narration/music were not embedded.
+The most common cause: audio sources were mixed externally but never passed in the Remotion
+`audio` prop. Fix: add `audio.narration` and `audio.music` to composition props and re-render.
+
+**Step 2: Extract review frames** at scene midpoints and visually inspect each one.
+
+**Step 3: Transcribe the rendered video's audio** using WhisperX/transcriber tool.
+- If 0 words returned → audio is silent despite stream existing → investigate
+- If word count < 80% of script → audio is cut off → investigate
+- Compare last transcribed word to last scripted word
+
+**Step 4: Present structured review** to user with file stats, audio verification results,
+visual findings, and caption status before declaring the video complete.
+
 ## Quality Checklist
 
 - [ ] Composition duration matches sum of scene durations minus transition overlaps
 - [ ] All `staticFile()` references resolve to existing assets
 - [ ] Transitions don't cut off content (account for overlap in timing)
+- [ ] **Audio stream present in rendered output** (ffprobe confirms codec_type: "audio")
+- [ ] **Narration words verified via transcription** (not just assumed from props)
 - [ ] Audio layers are in sync with visual scenes
+- [ ] Captions/subtitles rendering correctly (Remotion CaptionOverlay preferred over FFmpeg SRT)
 - [ ] Theme colors match the active style playbook
 - [ ] Output resolution and FPS match the target media profile
 - [ ] Render completes without Chromium timeout errors
 - [ ] Final output plays correctly on target platform
+- [ ] Text-bearing scenes (CTA, titles) use Remotion text_card, NOT AI-generated images with text
diff --git a/skills/creative/image-provider-usage.md b/skills/creative/image-provider-usage.md
index baed41f..179e64c 100644
--- a/skills/creative/image-provider-usage.md
+++ b/skills/creative/image-provider-usage.md
@@ -12,7 +12,7 @@
 | `flux_image` | FLUX 2 Pro via fal.ai | ~$0.03-0.05 | ~5-10s | Photorealism, general purpose, workhorse |
 | `grok_image` | Grok Imagine Image (xAI) | $0.02/output + $0.002/input edit image | ~5-15s | Image edits, style transfer, multi-image compositing |
 | `openai_image` | GPT Image 1 (OpenAI) | ~$0.01-0.17 | ~5-15s | Complex instructions, text in images, multi-element |
-| `recraft_image` | Recraft V4 via fal.ai | ~$0.04-0.25 | ~5-10s | Logos, SVG vectors, brand assets, text rendering |
+| `recraft_image` | Recraft V4 via fal.ai | ~$0.04-0.25 | ~5-10s | Logos, SVG vectors, brand assets, text rendering (see caveat below) |
 | `local_diffusion` | Stable Diffusion (local) | Free | ~30s+ | Offline, privacy, free |
 | `image_gen` | Multi (legacy, deprecated) | Varies | Varies | **Deprecated** — use `image_selector` or per-provider tools |
 
@@ -46,6 +46,12 @@
 | **Budget/free project** | `pexels_image` or `pixabay_image` | Free, immediate | `local_diffusion` |
 | **Offline/air-gapped** | `local_diffusion` | No network needed | — |
 
+## Provider-Specific Caveats
+
+### Recraft V4 via fal.ai
+- **`style` parameter causes 422 errors** (as of 2026-04). The `style` enum values (`digital_illustration`, `realistic_image`, etc.) are rejected by fal.ai's Recraft V4 endpoint. **Workaround:** encode style direction in the prompt text instead (e.g. "digital illustration of a tooth cross-section" rather than `style="digital_illustration"`). The `image_size` and `colors` parameters work fine.
+- **Text rendering is unreliable for exact business names.** Recraft (like all AI image models) may hallucinate wrong text. For any scene where text must be verbatim (CTA screens, business names, phone numbers), use Remotion `text_card` instead of generating an image with text.
+
 ## Cost-Quality Tradeoff
 
 ```
diff --git a/skills/pipelines/explainer/asset-director.md b/skills/pipelines/explainer/asset-director.md
index 0399df0..034acba 100644
--- a/skills/pipelines/explainer/asset-director.md
+++ b/skills/pipelines/explainer/asset-director.md
@@ -210,6 +210,7 @@ the AI model's training data — it may be wrong or outdated.
 - **Ignoring narration timing**: If TTS produces 12s of audio for a 10s section, the edit phase will struggle. Check durations.
 - **Missing pronunciation guide**: "PostgreSQL" or "Kubernetes" will be mispronounced without explicit guidance.
 - **One retry then give up**: If an image doesn't match, refine the prompt specifically — don't just retry the same prompt.
+- **AI-generating images with exact text (CTA, business names, contact info)**: AI image models frequently hallucinate wrong text — wrong business name, wrong phone number, misspelled words. **Never use AI image generation for scenes where text must be verbatim.** Use Remotion `text_card` type instead. This applies to: CTA screens, title cards with business names, contact info overlays, legal disclaimers. If a scene's `type` is `text_card` in the scene plan, do NOT generate an image for it — skip it and let the compose stage render it natively in Remotion.
 
 
 ## When You Do Not Know How
diff --git a/skills/pipelines/explainer/compose-director.md b/skills/pipelines/explainer/compose-director.md
index 1370e39..e3cadf6 100644
--- a/skills/pipelines/explainer/compose-director.md
+++ b/skills/pipelines/explainer/compose-director.md
@@ -22,19 +22,24 @@ This is the last technical stage before the video exists as a playable file. Eve
 
 Based on the edit decisions, pick the rendering approach:
 
-**FFmpeg pipeline** (simpler videos):
+**Remotion render** (DEFAULT — use this unless explicitly overridden):
+- Animated text cards, stat cards, chart scenes
+- Complex transitions (morph, zoom, ken-burns)
+- Programmatic motion graphics
+- Audio embedding (narration + music with fade/volume)
+- Word-level captions via CaptionOverlay component
+- Best for: ALL explainer videos, both image-based and animation-heavy
+
+**FFmpeg pipeline** (FALLBACK — only when Remotion is unavailable):
 - Static images with Ken Burns
 - Audio layering
-- Subtitle burn-in
-- Best for: diagram-heavy, image-based explainers
+- SRT subtitle burn-in
+- Best for: environments without Node.js/Remotion installed
 
-**Remotion render** (motion-heavy videos):
-- Animated text cards, stat cards
-- Complex transitions (morph, zoom)
-- Programmatic motion graphics
-- Best for: flat-motion-graphics playbook, animation-heavy plans
-
-You can combine both: Remotion for animated segments, FFmpeg for final assembly.
+**IMPORTANT: When using Remotion, ALL of these go through Remotion — not FFmpeg:**
+- Audio (narration + music) → Remotion `audio` prop, NOT external audio_mixer
+- Subtitles → Remotion `captions` prop (word-level), NOT SRT burn via FFmpeg
+- Text overlays (CTA, titles) → Remotion `text_card` cut type, NOT AI-generated images
 
 ### Step 2: Audio Acquisition (Narration, Music, Subtitles)
 
@@ -170,19 +175,34 @@ for the proven formula — especially the all-dark-background rule for visual co
 
 ### Step 5: Audio Post-Processing
 
+**Remotion path (DEFAULT):** Skip external audio mixing entirely. Remotion handles all audio
+natively via `<Audio>` components. Pass audio sources in the composition props:
+```json
+{
+  "audio": {
+    "narration": { "src": "project/narration.mp3", "volume": 1.0 },
+    "music": { "src": "project/music.mp3", "volume": 0.12, "fadeInSeconds": 1.5, "fadeOutSeconds": 2.5 }
+  }
+}
+```
+Remotion renders audio and video in a single pass — no external muxing needed.
+Do NOT use `audio_mixer` for ducking/mixing when rendering via Remotion.
+
+**FFmpeg fallback (ONLY when Remotion is unavailable):**
 Call the `audio_mixer` tool to:
 1. Layer narration segments in order
 2. Mix background music at playbook volume
 3. Apply ducking (music dips during narration)
 4. Normalize overall audio levels
 5. Output the final mixed audio track
-
 The video_compose tool will mux this with the video.
 
 ### Step 5b: Generate Subtitles (Mandatory)
 
 Subtitles are mandatory for all explainer content. Generate them from the narration audio — do NOT skip this step.
 
+**Remotion path (DEFAULT — when using Remotion render):**
+
 1. **Transcribe** the full narration using the `transcriber` tool (whisperx):
    ```python
    from tools.analysis.transcriber import Transcriber
@@ -195,39 +215,50 @@ Subtitles are mandatory for all explainer content. Generate them from the narrat
    # result.data contains segments with word-level timestamps
    ```
 
-2. **Generate SRT** from the transcription using `subtitle_gen`:
+2. **Convert to Remotion WordCaption format** (NOT SRT):
+   ```python
+   captions = []
+   for segment in result.data['segments']:
+       for word_info in segment.get('words', []):
+           captions.append({
+               'word': word_info['word'],
+               'startMs': int(word_info['start'] * 1000),
+               'endMs': int(word_info['end'] * 1000),
+           })
+   ```
+
+3. **Add captions to composition props** — they go in the `captions` array alongside `cuts` and `audio`:
+   ```json
+   {
+     "cuts": [...],
+     "audio": {...},
+     "captions": [
+       { "word": "Root", "startMs": 120, "endMs": 340 },
+       { "word": "canals", "startMs": 340, "endMs": 680 }
+     ]
+   }
+   ```
+
+   Remotion's CaptionOverlay renders these as word-by-word highlighted captions with the theme's
+   `captionHighlightColor` and `captionBackgroundColor`. This is superior to FFmpeg SRT burn because
+   it produces animated word-level highlighting synchronized to narration.
+
+**FFmpeg fallback (ONLY when Remotion is unavailable):**
+
+If Remotion is not available, fall back to SRT generation + FFmpeg burn:
    ```python
    from tools.subtitle.subtitle_gen import SubtitleGen
-   result = SubtitleGen().execute({
+   SubtitleGen().execute({
        'segments': transcription_data['segments'],
        'format': 'srt',
        'output_path': 'projects/<project>/assets/subtitles.srt',
        'max_words_per_cue': 8,
        'max_chars_per_line': 42
    })
+   # Then burn with video_compose operation='burn_subtitles'
    ```
 
-3. **Burn subtitles** into the video using `video_compose`:
-   ```python
-   from tools.video.video_compose import VideoCompose
-   result = VideoCompose().execute({
-       'operation': 'burn_subtitles',
-       'input_path': 'projects/<project>/renders/output.mp4',
-       'output_path': 'projects/<project>/renders/final.mp4',
-       'subtitle_path': 'projects/<project>/assets/subtitles.srt',
-       'subtitle_style': {
-           'font': '<from playbook typography.headings.font or Arial>',
-           'font_size': 22,
-           'primary_color': '&HFFFFFF',
-           'outline_color': '&H000000',
-           'outline_width': 2,
-           'margin_v': 50,
-           'alignment': 2
-       }
-   })
-   ```
-
-**The final deliverable is the subtitled version**, not the pre-subtitle render.
+**The final deliverable MUST have subtitles** — either via Remotion captions or FFmpeg burn.
 
 ### Step 5c: Pre-Render Validation (Mandatory)
 
@@ -250,14 +281,30 @@ Common catches:
 
 **Do not skip this step.** If validation fails, fix the issue and re-validate before rendering.
 
-### Step 6: Post-Render Self-Review (Mandatory)
+### Step 6: Post-Render Self-Review (Mandatory — ALL steps required)
 
 After rendering, the agent **must review its own output** before presenting to the user. This catches issues the validator can't see (visual quality, audio sync, subtitle readability).
 
-**6a. Extract review frames:**
+**CRITICAL: You MUST complete ALL of steps 6a through 6e. Do NOT skip any step.
+The most common agent failure is doing 6a (frames) and 6c (visual) while skipping
+6b (audio transcription) — which misses catastrophic issues like missing audio entirely.**
+
+**6a. Probe rendered file (FIRST — gate for all other checks):**
+```bash
+ffprobe -v quiet -print_format json -show_format -show_streams rendered_video.mp4
+```
+Verify:
+- Video stream exists (codec_type: "video") with correct resolution
+- **Audio stream exists (codec_type: "audio")** — if NO audio stream, STOP and fix immediately
+- Duration is within ±5% of target
+- File size is reasonable (not 0 bytes)
+
+**If audio stream is missing: the render did not embed audio. Do NOT proceed to present
+the video to the user. Fix the audio configuration and re-render.**
+
+**6b. Extract review frames:**
 ```python
 from tools.analysis.frame_sampler import FrameSampler
-# Extract one frame per scene at the midpoint
 midpoints = [(cut['in_seconds'] + cut['out_seconds']) / 2 for cut in cuts]
 FrameSampler().execute({
     'input_path': 'path/to/rendered_video.mp4',
@@ -268,37 +315,41 @@ FrameSampler().execute({
 })
 ```
 
-**6b. Transcribe rendered audio:**
+**6c. Transcribe rendered audio (MANDATORY — do NOT skip):**
 ```python
 from tools.analysis.transcriber import Transcriber
-Transcriber().execute({
+result = Transcriber().execute({
     'input_path': 'path/to/rendered_video.mp4',
     'model_size': 'base',
     'language': 'en',
     'output_dir': 'path/to/review-frames',
 })
-# Verify all narration words are present and not cut off
+# If result returns 0 words: audio is silent/missing — STOP and fix
+# If word count < 80% of script word count: audio is cut off — investigate
 ```
 
-**6c. Visual inspection — review each frame:**
+**6d. Visual inspection — review each frame:**
 - Does the background color/gradient match intent? (watch for white backgrounds on dark-themed videos)
 - Are images rendering correctly? (not blank, not stretched)
-- Are subtitles visible and properly spaced?
+- Are subtitles/captions visible and properly spaced?
 - Are overlays (section titles, stat reveals) positioned correctly?
 - Is the opening scene visually strong? (important for social media thumbnails)
+- Does the CTA/closing screen show correct text? (AI-generated text in images frequently hallucinates — use Remotion text_card for any text that must be exact)
 
-**6d. Audio inspection — check transcript:**
+**6e. Audio inspection — check transcript against script:**
 - Is the full narration captured? (compare last transcribed word to last scripted word)
 - Any words cut off at the end? (narration exceeding video duration)
 - Timing alignment — do narration segments roughly match their intended scenes?
+- Is background music audible? (transcriber may not capture music, but ffprobe confirms audio stream)
 
-**6e. Compile and present review to user:**
+**6f. Compile and present review to user:**
 
 > **Post-render review for "[Video Title]":**
 >
-> **Audio:** [Complete/Cut off at Xs] — all N words captured / last sentence missing
+> **File:** [duration]s, [resolution], [file size] — audio stream: [present/MISSING]
+> **Audio:** [Complete/Cut off at Xs] — [N]/[M] words transcribed from rendered output
 > **Visuals:** [N scenes inspected] — [issues or "all scenes rendering correctly"]
-> **Subtitles:** [Present/Missing] — [spacing ok / words running together]
+> **Captions:** [Remotion CaptionOverlay / FFmpeg SRT / MISSING] — [word-level highlight working / issues]
 > **Issues found:** [list any issues with severity]
 >
 > **Recommendations:** [what to fix, if anything]
diff --git a/skills/pipelines/explainer/scene-director.md b/skills/pipelines/explainer/scene-director.md
index d9512ed..86787aa 100644
--- a/skills/pipelines/explainer/scene-director.md
+++ b/skills/pipelines/explainer/scene-director.md
@@ -223,3 +223,4 @@ Call `handle_explainer_scene_plan(state, {"scene_plan": scene_plan_json})` to va
 - **Vague required_assets**: "An image about databases" is useless for prompt engineering. "Isometric illustration of a vector database with embedding vectors floating in 3D space, using the playbook's blue-green palette" is actionable.
 - **Preset thinking**: A scene plan that says "make it flat-motion-graphics" is not enough. The planner must specify what makes THIS video's motion graphics feel distinct.
 - **Static scenes for dynamic concepts**: If the narrator describes a process or transformation, the visual should move. Use animation or progressive reveal, not a static image.
+- **Using `generated` type for CTA/closing screens with exact text**: AI image models hallucinate text — wrong business names, misspelled words, wrong phone numbers. Any scene with verbatim text (CTA, business info, contact details, legal) MUST be `type: "text_card"` so Remotion renders the text exactly. Never plan a `generated` image for a scene where text accuracy matters.
diff --git a/tools/graphics/recraft_image.py b/tools/graphics/recraft_image.py
index 4e9e364..afc312c 100644
--- a/tools/graphics/recraft_image.py
+++ b/tools/graphics/recraft_image.py
@@ -144,6 +144,14 @@ class RecraftImage(BaseTool):
         if inputs.get("image_size"):
             payload["image_size"] = inputs["image_size"]
         if inputs.get("style"):
+            # NOTE: As of 2026-04, fal.ai's Recraft V4 endpoint rejects the
+            # `style` parameter with a 422 Unprocessable Entity error. The
+            # style enum values (digital_illustration, realistic_image, etc.)
+            # are NOT accepted by the /fal-ai/recraft/v4/text-to-image route.
+            # Workaround: encode the style direction in the prompt text instead
+            # (e.g. "digital illustration of..." rather than style="digital_illustration").
+            # We still pass the parameter through in case fal.ai re-enables it,
+            # but callers should be aware this may fail.
             payload["style"] = inputs["style"]
         if inputs.get("colors"):
             payload["colors"] = inputs["colors"]