diff --git a/.gsd/milestones/M024/M024-ROADMAP.md b/.gsd/milestones/M024/M024-ROADMAP.md
index 49536c7..9a7efba 100644
--- a/.gsd/milestones/M024/M024-ROADMAP.md
+++ b/.gsd/milestones/M024/M024-ROADMAP.md
@@ -8,7 +8,7 @@ Shorts pipeline goes end-to-end with captioning and templates. Player gets key m
 |----|-------|------|---------|------|------------|
 | S01 | [A] Shorts Publishing Flow | medium | — | ✅ | Creator approves a short → it renders → gets a shareable URL and embed code |
 | S02 | [A] Key Moment Pins on Player Timeline | low | — | ✅ | Key technique moments appear as clickable pins on the player timeline |
-| S03 | [A] Embed Support (iframe Snippet) | low | — | ⬜ | Creators can copy an iframe embed snippet to put the player on their own site |
+| S03 | [A] Embed Support (iframe Snippet) | low | — | ✅ | Creators can copy an iframe embed snippet to put the player on their own site |
 | S04 | [B] Auto-Captioning + Template System | medium | — | ⬜ | Shorts have Whisper-generated animated subtitles and creator-configurable intro/outro cards |
 | S05 | [B] Citation UX Improvements | low | — | ⬜ | Chat citations show timestamp links that seek the player and source cards with video thumbnails |
 | S06 | Forgejo KB Update — Shorts, Embed, Citations | low | S01, S02, S03, S04, S05 | ⬜ | Forgejo wiki updated with shorts pipeline, embed system, citation architecture |
diff --git a/.gsd/milestones/M024/slices/S03/S03-SUMMARY.md b/.gsd/milestones/M024/slices/S03/S03-SUMMARY.md
new file mode 100644
index 0000000..12b77a5
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S03/S03-SUMMARY.md
@@ -0,0 +1,95 @@
+---
+id: S03
+parent: M024
+milestone: M024
+provides:
+  - EmbedPlayer page at /embed/:videoId
+  - Shared copyToClipboard utility
+  - Copy Embed Code button on WatchPage
+requires:
+  []
+affects:
+  - S06
+key_files:
+  - frontend/src/utils/clipboard.ts
+  - frontend/src/pages/EmbedPlayer.tsx
+  - frontend/src/pages/EmbedPlayer.module.css
+  - frontend/src/pages/ShortPlayer.tsx
+  - frontend/src/App.tsx
+  - frontend/src/pages/WatchPage.tsx
+  - frontend/src/App.css
+key_decisions:
+  - Embed route rendered at top-level Routes before AppShell fallback for chrome-free iframe rendering
+  - Audio-only embeds use height 120 vs 405 for video in generated snippet
+  - Branding link opens origin in new tab with noopener for iframe safety
+  - copyToClipboard extracted to shared utility for reuse across ShortPlayer and WatchPage
+patterns_established:
+  - Top-level Routes in App.tsx for chrome-free pages that skip AppShell (header/nav/footer)
+observability_surfaces:
+  - none
+drill_down_paths:
+  - .gsd/milestones/M024/slices/S03/tasks/T01-SUMMARY.md
+  - .gsd/milestones/M024/slices/S03/tasks/T02-SUMMARY.md
+duration: ""
+verification_result: passed
+completed_at: 2026-04-04T11:00:25.948Z
+blocker_discovered: false
+---
+
+# S03: [A] Embed Support (iframe Snippet)
+
+**Creators can copy an iframe embed snippet from WatchPage, and /embed/:videoId renders a chrome-free player suitable for iframe embedding.**
+
+## What Happened
+
+Built the embed support feature in two tasks. T01 extracted copyToClipboard from ShortPlayer into a shared utility at `frontend/src/utils/clipboard.ts`, then created the EmbedPlayer page (`EmbedPlayer.tsx` + `EmbedPlayer.module.css`) — a full-viewport dark-background player that fetches video detail and renders either VideoPlayer or AudioWaveform based on content type, with a small "Powered by Chrysopedia" branding link.
+
+T02 wired the `/embed/:videoId` route at the top level of App.tsx's Routes (before the AppShell catch-all), so embed pages render without header/nav/footer. Added a "Copy Embed Code" button to WatchPage's header that generates an iframe snippet with audio-aware height (120px for audio-only, 405px for video) and shows 2-second "Copied!" feedback. The EmbedPlayer chunk is code-split via React.lazy.
+
+Both `tsc --noEmit` and `npm run build` pass cleanly. The embed route is isolated from the app shell, the clipboard utility is shared, and the iframe snippet includes correct dimensions per content type.
+
+## Verification
+
+- `cd frontend && npx tsc --noEmit` — exit 0, zero type errors
+- `cd frontend && npm run build` — exit 0, EmbedPlayer code-split into own chunk, 190 modules transformed
+- Files confirmed: clipboard.ts, EmbedPlayer.tsx, EmbedPlayer.module.css all present
+- App.tsx has /embed/:videoId route before AppShell catch-all (line 233)
+- WatchPage.tsx generates iframe snippet with audio-aware height (line 35)
+
+## Requirements Advanced
+
+None.
+
+## Requirements Validated
+
+None.
+
+## New Requirements Surfaced
+
+None.
+
+## Requirements Invalidated or Re-scoped
+
+None.
+
+## Deviations
+
+Added .watch-page__header-top flex container in WatchPage for title/button layout — minor structural addition not in plan.
+
+## Known Limitations
+
+None.
+
+## Follow-ups
+
+None.
+
+## Files Created/Modified
+
+- `frontend/src/utils/clipboard.ts` — New shared copyToClipboard utility extracted from ShortPlayer
+- `frontend/src/pages/EmbedPlayer.tsx` — New chrome-free embed player page with video/audio support
+- `frontend/src/pages/EmbedPlayer.module.css` — Full-viewport dark layout styles for embed player
+- `frontend/src/pages/ShortPlayer.tsx` — Updated to import copyToClipboard from shared utility
+- `frontend/src/App.tsx` — Added /embed/:videoId route before AppShell catch-all
+- `frontend/src/pages/WatchPage.tsx` — Added Copy Embed Code button with audio-aware iframe snippet
+- `frontend/src/App.css` — Added styles for embed copy button
diff --git a/.gsd/milestones/M024/slices/S03/S03-UAT.md b/.gsd/milestones/M024/slices/S03/S03-UAT.md
new file mode 100644
index 0000000..181731d
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S03/S03-UAT.md
@@ -0,0 +1,52 @@
+# S03: [A] Embed Support (iframe Snippet) — UAT
+
+**Milestone:** M024
+**Written:** 2026-04-04T11:00:25.948Z
+
+## UAT: Embed Support (iframe Snippet)
+
+### Preconditions
+- Chrysopedia frontend running (ub01:8096 or local dev server)
+- At least one video-type and one audio-only source video in the database
+
+### Test 1: Copy Embed Code — Video Content
+1. Navigate to a WatchPage for a video that has `video_url` (e.g., `/watch/{videoId}`)
+2. Locate the "Copy Embed Code" button in the page header
+3. Click the button
+4. **Expected:** Button text changes to "Copied!" for ~2 seconds, then reverts to "Copy Embed Code"
+5. Paste clipboard contents into a text editor
+6. **Expected:** Clipboard contains `<iframe src="http://{origin}/embed/{videoId}" width="720" height="405" frameborder="0" allowfullscreen></iframe>`
+
+### Test 2: Copy Embed Code — Audio-Only Content
+1. Navigate to a WatchPage for an audio-only source video (no `video_url`)
+2. Click "Copy Embed Code"
+3. Paste clipboard contents
+4. **Expected:** iframe height is `120` (not 405): `<iframe src="..." width="720" height="120" ...></iframe>`
+
+### Test 3: Embed Route — Chrome-Free Video
+1. Navigate directly to `/embed/{videoId}` for a video source
+2. **Expected:** Full-viewport dark background, video player fills the space, no site header/nav/footer
+3. **Expected:** Small "Powered by Chrysopedia" link at the bottom
+4. **Expected:** Player controls (play/pause, seek, volume) are functional
+
+### Test 4: Embed Route — Chrome-Free Audio
+1. Navigate to `/embed/{videoId}` for an audio-only source
+2. **Expected:** Audio waveform or audio player renders instead of video player
+3. **Expected:** Same chrome-free layout with branding link
+
+### Test 5: Embed Route with Start Time
+1. Navigate to `/embed/{videoId}?t=30`
+2. **Expected:** Player starts at or seeks to the 30-second mark
+
+### Test 6: Embed Route — Invalid Video ID
+1. Navigate to `/embed/nonexistent-id`
+2. **Expected:** Error state displayed (not a blank page or crash)
+
+### Test 7: iframe Integration
+1. Create a local HTML file with the copied iframe snippet
+2. Open it in a browser
+3. **Expected:** Chrysopedia player loads inside the iframe, video plays, no app chrome visible
+
+### Edge Cases
+- **Rapid clicks on Copy Embed Code:** Should not stack timeouts or cause flickering — button stays "Copied!" and timer resets
+- **Narrow viewport in iframe:** Embed player should be responsive, scaling to container width
diff --git a/.gsd/milestones/M024/slices/S03/tasks/T02-VERIFY.json b/.gsd/milestones/M024/slices/S03/tasks/T02-VERIFY.json
new file mode 100644
index 0000000..ca92f5c
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S03/tasks/T02-VERIFY.json
@@ -0,0 +1,16 @@
+{
+  "schemaVersion": 1,
+  "taskId": "T02",
+  "unitId": "M024/S03/T02",
+  "timestamp": 1775300354664,
+  "passed": true,
+  "discoverySource": "task-plan",
+  "checks": [
+    {
+      "command": "cd frontend",
+      "exitCode": 0,
+      "durationMs": 14,
+      "verdict": "pass"
+    }
+  ]
+}
diff --git a/.gsd/milestones/M024/slices/S04/S04-PLAN.md b/.gsd/milestones/M024/slices/S04/S04-PLAN.md
index 731e170..f2a9b2c 100644
--- a/.gsd/milestones/M024/slices/S04/S04-PLAN.md
+++ b/.gsd/milestones/M024/slices/S04/S04-PLAN.md
@@ -1,6 +1,75 @@
 # S04: [B] Auto-Captioning + Template System
 
-**Goal:** Add auto-captioning and template system to shorts pipeline
+**Goal:** Shorts have Whisper-generated animated subtitles and creator-configurable intro/outro cards
 **Demo:** After this: Shorts have Whisper-generated animated subtitles and creator-configurable intro/outro cards
 
 ## Tasks
+- [x] **T01: Created ASS subtitle generator with karaoke word-by-word highlighting and wired it into the shorts generation stage with non-blocking caption enrichment** — Create `caption_generator.py` that converts word-level timings into ASS (Advanced SubStation Alpha) subtitle format with word-by-word karaoke highlighting. Modify `shorts_generator.py` to accept an optional ASS file path and chain the `ass=` filter into the ffmpeg `-vf` string. Wire transcript loading and caption generation into `stage_generate_shorts` in `stages.py`. Add `captions_enabled` boolean column to `GeneratedShort` model. Write unit tests for caption generation.
+
+Steps:
+1. Read `backend/pipeline/highlight_scorer.py` for `extract_word_timings` signature and output format. Read `backend/pipeline/shorts_generator.py` for `extract_clip` and `PRESETS`. Read `backend/pipeline/stages.py:2869-2990` for `stage_generate_shorts` flow.
+2. Create `backend/pipeline/caption_generator.py`:
+   - `generate_ass_captions(word_timings: list[dict], clip_start: float, style_config: dict | None = None) -> str` — returns ASS file content as string. Each word gets a `Dialogue` line. Use `{\k}` karaoke tags for word-by-word highlight timing. Style: bold white text, centered bottom 15%, black outline. Offset all word times by `-clip_start` to make them clip-relative.
+   - `write_ass_file(ass_content: str, output_path: Path) -> Path` — writes to disk, returns path.
+3. Modify `extract_clip()` in `shorts_generator.py`: add optional `ass_path: Path | None = None` parameter. When provided, append `,ass={ass_path}` to the `vf_filter` string before passing to ffmpeg. Ensure the ASS filter comes after scale/pad filters.
+4. Add `captions_enabled: Mapped[bool] = mapped_column(default=False, server_default='false')` to `GeneratedShort` in `models.py`.
+5. Create Alembic migration `027_add_captions_enabled.py` for the new column.
+6. Modify `stage_generate_shorts` in `stages.py`:
+   - After loading the highlight, load `source_video.transcript_path` and parse transcript JSON (reuse the pattern from line ~2465).
+   - Call `extract_word_timings(transcript_data, clip_start, clip_end)` to get word timings for the clip window.
+   - If word timings are non-empty, call `generate_ass_captions()` and `write_ass_file()` to a temp path.
+   - Pass the ASS path to `extract_clip()`. Set `short.captions_enabled = True`.
+   - If word timings are empty, log a warning and proceed without captions.
+7. Create `backend/pipeline/test_caption_generator.py` with tests:
+   - Valid word timings → correct ASS output with proper timing math
+   - Empty word timings → empty ASS (or raise, depending on design)
+   - Clip offset applied correctly (word at t=10.5 with clip_start=10.0 becomes t=0.5)
+   - ASS format structure (header, style block, dialogue lines)
+8. Run tests: `cd backend && python -m pytest pipeline/test_caption_generator.py -v`
+  - Estimate: 2h
+  - Files: backend/pipeline/caption_generator.py, backend/pipeline/shorts_generator.py, backend/pipeline/stages.py, backend/models.py, alembic/versions/027_add_captions_enabled.py, backend/pipeline/test_caption_generator.py
+  - Verify: cd backend && python -m pytest pipeline/test_caption_generator.py -v && python -c "from pipeline.caption_generator import generate_ass_captions; print('import ok')"
+- [ ] **T02: Build card renderer and concat pipeline for intro/outro templates** — Create `card_renderer.py` that generates intro/outro card video segments using ffmpeg lavfi (color + drawtext). Add `shorts_template` JSONB column to Creator model. Implement ffmpeg concat demuxer logic to assemble intro + main clip + outro into final short. Wire into `stage_generate_shorts`. Write unit tests for card renderer.
+
+Steps:
+1. Read T01 outputs: `backend/pipeline/caption_generator.py`, modified `shorts_generator.py` and `stages.py`.
+2. Add `shorts_template: Mapped[dict | None] = mapped_column(JSONB, nullable=True)` to `Creator` model in `models.py`. Create Alembic migration `028_add_shorts_template.py`.
+3. Create `backend/pipeline/card_renderer.py`:
+   - `render_card(text: str, duration_secs: float, width: int, height: int, accent_color: str = '#22d3ee', font_family: str = 'Inter') -> list[str]` — returns ffmpeg command args that generate a card mp4 from lavfi input (`color=c=black:s={w}x{h}:d={dur}` with `drawtext` for centered text, accent color underline/glow).
+   - `render_card_to_file(text: str, duration_secs: float, width: int, height: int, output_path: Path, accent_color: str = '#22d3ee', font_family: str = 'Inter') -> Path` — executes the ffmpeg command, returns output path.
+   - `concat_segments(segments: list[Path], output_path: Path) -> Path` — writes a concat demuxer list file, runs `ffmpeg -f concat -safe 0 -i list.txt -c copy output.mp4`, returns output path. All segments must share codec settings.
+4. Modify `shorts_generator.py`: add `extract_clip_with_template(input_path, output_path, start_secs, end_secs, vf_filter, ass_path=None, intro_path=None, outro_path=None) -> None` that extracts the main clip (with optional captions), then if intro/outro paths are provided, concats them via `concat_segments()`.
+5. Modify `stage_generate_shorts` in `stages.py`:
+   - After loading highlight, also load `highlight.source_video.creator` to access `creator.shorts_template`.
+   - If `shorts_template` exists and `show_intro` is true, call `render_card_to_file()` for intro. Same for outro.
+   - Pass intro/outro paths to the clip extraction. Use codec-compatible settings (libx264, aac, same resolution from preset spec).
+   - If no template, proceed without cards (existing behavior preserved).
+6. Create `backend/pipeline/test_card_renderer.py` with tests:
+   - `render_card()` returns valid ffmpeg command with correct dimensions and duration
+   - `concat_segments()` generates correct concat list file content
+   - Template config parsing handles missing/partial fields with defaults
+7. Run tests: `cd backend && python -m pytest pipeline/test_card_renderer.py -v`
+  - Estimate: 2h
+  - Files: backend/pipeline/card_renderer.py, backend/pipeline/shorts_generator.py, backend/pipeline/stages.py, backend/models.py, alembic/versions/028_add_shorts_template.py, backend/pipeline/test_card_renderer.py
+  - Verify: cd backend && python -m pytest pipeline/test_card_renderer.py -v && python -c "from pipeline.card_renderer import render_card, concat_segments; print('import ok')"
+- [ ] **T03: Template API endpoints and frontend template config UI** — Add REST endpoints for reading and updating creator shorts template config. Add template configuration UI to the HighlightQueue page — color picker, text inputs, duration controls, and intro/outro toggles. Add a caption toggle to the short generation flow.
+
+Steps:
+1. Read T02 outputs to understand the shorts_template schema on the Creator model.
+2. Create or extend `backend/routers/creators.py` with two endpoints:
+   - `GET /api/v1/admin/creators/{creator_id}/shorts-template` — returns current `shorts_template` JSONB or default config if null.
+   - `PUT /api/v1/admin/creators/{creator_id}/shorts-template` — validates and saves template config. Pydantic schema: `ShortsTemplateUpdate` with fields: intro_text (str, max 100), outro_text (str, max 100), accent_color (str, hex pattern), font_family (str), intro_duration_secs (float, 1.0-5.0), outro_duration_secs (float, 1.0-5.0), show_intro (bool), show_outro (bool).
+   - Both endpoints require admin auth.
+3. Add `ShortsTemplateConfig` and `ShortsTemplateUpdate` Pydantic schemas to `backend/schemas.py`.
+4. Create `frontend/src/api/templates.ts` — API client functions: `fetchShortsTemplate(creatorId)`, `updateShortsTemplate(creatorId, config)`.
+5. Add template config UI to `frontend/src/pages/HighlightQueue.tsx`:
+   - A collapsible "Shorts Template" section in the sidebar or above the queue.
+   - Fields: intro text, outro text, accent color (HTML color input), intro/outro duration sliders (1-5s), show intro/outro toggles.
+   - Save button that calls `updateShortsTemplate()`.
+   - Load current template on mount when a creator is selected.
+6. Add a "Captions" toggle checkbox to the short generation trigger in HighlightQueue — when unchecked, pass `captions=false` query param to the generate endpoint. Update the `POST /api/v1/admin/highlights/{id}/generate-shorts` handler (in `backend/routers/creator_highlights.py` or similar) to accept and forward an optional `captions` param.
+7. Verify frontend builds: `cd frontend && npm run build`
+8. Verify API imports: `cd backend && python -c "from routers.creators import router; print('ok')"`
+  - Estimate: 2h
+  - Files: backend/routers/creators.py, backend/schemas.py, frontend/src/api/templates.ts, frontend/src/pages/HighlightQueue.tsx, frontend/src/pages/HighlightQueue.module.css, backend/routers/creator_highlights.py
+  - Verify: cd frontend && npm run build 2>&1 | tail -5 && cd ../backend && python -c "from routers.creators import router; print('ok')"
diff --git a/.gsd/milestones/M024/slices/S04/S04-RESEARCH.md b/.gsd/milestones/M024/slices/S04/S04-RESEARCH.md
new file mode 100644
index 0000000..4f8aeb3
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S04/S04-RESEARCH.md
@@ -0,0 +1,108 @@
+# S04 Research: Auto-Captioning + Template System
+
+## Summary
+
+Add Whisper-derived animated subtitles to generated shorts and a creator-configurable intro/outro card system. Medium complexity — captioning uses established ASS format burned via ffmpeg (well-documented pattern), template system needs schema + UI + ffmpeg concat.
+
+## Recommendation
+
+Build captioning first (riskiest, validates the word-timing → ASS → ffmpeg pipeline), then template system (straightforward concat). Keep both features optional per-short — existing shorts should still generate without captions/templates if data is missing.
+
+## Implementation Landscape
+
+### Current State
+
+1. **Shorts generation pipeline** (`backend/pipeline/stages.py:2869–3055`): `stage_generate_shorts` Celery task extracts clips via `extract_clip()` from `shorts_generator.py`, uploads to MinIO, creates `GeneratedShort` DB rows. Currently uses a simple `-vf` filter for scaling/padding only.
+
+2. **Word-level timing data**: Whisper transcripts stored on disk at `settings.transcript_storage_path` contain per-word `{word, start, end}` dicts inside each segment. The function `extract_word_timings()` in `highlight_scorer.py:186` already extracts words for a `[start_time, end_time]` window — this is the exact input needed for subtitle generation.
+
+3. **Transcript loading pattern** (`stages.py:2465–2485`): Stage loads transcript JSON from `source_video.transcript_path`, parses segments. Same pattern reusable in `stage_generate_shorts`.
+
+4. **Creator model** (`models.py:122`): Has `personality_profile` JSONB but no template/branding fields. No intro/outro configuration exists.
+
+5. **GeneratedShort model** (`models.py:836`): Has format_preset, minio_object_key, dimensions, status. No caption or template metadata columns.
+
+6. **Frontend**: `ShortPlayer.tsx` is a basic `<video>` player. `HighlightQueue.tsx` has short generation trigger UI. Neither has caption/template controls.
+
+### What Needs Building
+
+#### Part 1: Auto-Captioning
+
+- **ASS subtitle generator** (`backend/pipeline/caption_generator.py` — new): Takes word timings + clip offset → generates ASS (Advanced SubStation Alpha) format file with animated word-by-word highlighting. ASS chosen over SRT because it supports styling (font, color, position, animation) needed for social-media-style captions.
+- **ffmpeg filter chain update** (`shorts_generator.py`): Add `ass=` filter to the existing `-vf` chain. Current filter is e.g. `scale=1080:-2,pad=1080:1920:(ow-iw)/2:(oh-ih)/2:black` → becomes `scale=1080:-2,pad=1080:1920:(ow-iw)/2:(oh-ih)/2:black,ass=/tmp/captions.ass`.
+- **Transcript loading in stage_generate_shorts**: Reuse the transcript loading pattern from highlight detection stage. Load transcript JSON, call `extract_word_timings(data, clip_start, clip_end)` to get word-level timing for the clip window. Offset all word times by `-clip_start` so they're relative to clip start.
+- **No new DB columns needed for basic captioning** — captions are burned into the video file. Optional: add `captions_enabled: bool` to GeneratedShort for tracking.
+
+#### Part 2: Template System (Intro/Outro Cards)
+
+- **DB schema**: Add `shorts_template` JSONB column to `Creator` model (new Alembic migration). Schema:
+  ```json
+  {
+    "intro_text": "CreatorName presents",
+    "outro_text": "Follow for more",
+    "accent_color": "#22d3ee",
+    "font_family": "Inter",
+    "intro_duration_secs": 2.0,
+    "outro_duration_secs": 3.0,
+    "show_intro": true,
+    "show_outro": true
+  }
+  ```
+- **Card renderer** (`backend/pipeline/card_renderer.py` — new): Generate intro/outro card video segments using ffmpeg's `lavfi` input (color + drawtext filter). Each card is a short standalone mp4 clip. Then concat all three pieces (intro + main clip + outro) using ffmpeg concat demuxer.
+- **API endpoint**: `PUT /api/v1/admin/creators/{creator_id}/shorts-template` to save template config. `GET` to retrieve current config.
+- **Frontend**: Template config UI on the creator dashboard / highlight queue. Color picker, text inputs, toggle for intro/outro.
+
+### Key Technical Details
+
+**ASS subtitle format for word-by-word animation**: Each word gets its own dialogue line with `{\kN}` karaoke timing tags. The `\k` tag specifies duration in centiseconds before the next word highlights. Style block defines font (bold, white, centered at bottom 15%), with `\3c` for outline color. ffmpeg's `ass` filter (part of libass) renders this natively.
+
+**Time offset for captions**: Word timings from Whisper are absolute (relative to video start). For a clip starting at `clip_start`, subtract `clip_start` from each word's start/end to get clip-relative times. If intro card is prepended, further offset by `intro_duration_secs`.
+
+**ffmpeg concat approach for templates**: Use concat demuxer (`-f concat -safe 0 -i list.txt`) with a text file listing intro.mp4, main.mp4, outro.mp4. All three must share codec settings (h264, aac, same resolution). The card renderer must output clips matching the preset dimensions.
+
+**Fallback behavior**: If transcript has no word-level timings for the clip window (empty `words` arrays), skip captioning and generate without subtitles. If creator has no `shorts_template`, skip intro/outro. Both features are additive — the base clip extraction path is unchanged.
+
+### File Inventory
+
+| File | Role | Changes |
+|------|------|---------|
+| `backend/pipeline/caption_generator.py` | NEW — ASS subtitle generation from word timings |  |
+| `backend/pipeline/card_renderer.py` | NEW — ffmpeg-based intro/outro card generation |  |
+| `backend/pipeline/shorts_generator.py` | Modify `extract_clip()` or add `extract_clip_with_captions()` to chain ASS filter |  |
+| `backend/pipeline/stages.py` | Modify `stage_generate_shorts` to load transcript, generate captions, handle templates |  |
+| `backend/models.py` | Add `shorts_template` JSONB to Creator, optionally `captions_enabled` to GeneratedShort |  |
+| `alembic/versions/027_*.py` | NEW — migration for shorts_template column |  |
+| `backend/routers/creators.py` or new router | Template config GET/PUT endpoints |  |
+| `frontend/src/pages/HighlightQueue.tsx` | Caption toggle, template preview |  |
+| `frontend/src/api/creators.ts` or similar | API client for template config |  |
+
+### Natural Task Seams
+
+1. **Caption generator + ffmpeg integration** (riskiest — validate ASS generation and ffmpeg burning works): Build `caption_generator.py`, modify `shorts_generator.py` to accept optional ASS path, wire into `stage_generate_shorts` with transcript loading. Test with a real clip.
+
+2. **Template schema + card renderer** (medium risk — ffmpeg concat is well-documented but needs resolution matching): DB migration, `card_renderer.py`, concat logic in stage.
+
+3. **Template API + frontend** (low risk — standard CRUD + form UI): REST endpoints, creator dashboard UI for template configuration.
+
+### Verification Strategy
+
+- Unit test `caption_generator.py`: given word timings + offset → produces valid ASS content with correct timing math
+- Unit test `card_renderer.py`: given template config + preset → produces correct ffmpeg command
+- Integration: `stage_generate_shorts` with mock transcript data → verify ASS file created, ffmpeg called with subtitle filter
+- Frontend build passes with zero TypeScript errors
+- Manual: generate a short for a highlight with word-level timing data → verify captions appear in output video
+
+### Risks
+
+1. **ffmpeg ASS filter requires libass**: The Docker image must have libass compiled into ffmpeg. Standard `ffmpeg` packages on Debian/Ubuntu include it, but verify with `ffmpeg -filters | grep ass` inside the container.
+2. **Word timing gaps**: Some Whisper transcripts may have segments without word-level timings (older transcriptions). Fallback to segment-level timing (less granular but still useful) or skip captions entirely.
+3. **Concat codec mismatch**: Intro/outro cards must exactly match the main clip's codec settings. Use identical ffmpeg encoding params for all three segments.
+
+### Don't Hand-Roll
+
+- **ASS format**: Use the standard ASS spec — don't invent a custom subtitle format. ASS is the native format for ffmpeg's `ass` filter.
+- **Video concatenation**: Use ffmpeg's concat demuxer, not manual frame-by-frame stitching or Python video libraries.
+
+### Skill Discovery
+
+No additional professional agent skills needed. The work is Python + ffmpeg (both well-understood in this codebase) and standard React form UI. The `whisper` skill in available_skills is for transcription, not subtitle generation — not relevant here since we're consuming existing transcript data.
diff --git a/.gsd/milestones/M024/slices/S04/tasks/T01-PLAN.md b/.gsd/milestones/M024/slices/S04/tasks/T01-PLAN.md
new file mode 100644
index 0000000..2ce4baf
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S04/tasks/T01-PLAN.md
@@ -0,0 +1,50 @@
+---
+estimated_steps: 21
+estimated_files: 6
+skills_used: []
+---
+
+# T01: Build caption generator and wire into shorts pipeline
+
+Create `caption_generator.py` that converts word-level timings into ASS (Advanced SubStation Alpha) subtitle format with word-by-word karaoke highlighting. Modify `shorts_generator.py` to accept an optional ASS file path and chain the `ass=` filter into the ffmpeg `-vf` string. Wire transcript loading and caption generation into `stage_generate_shorts` in `stages.py`. Add `captions_enabled` boolean column to `GeneratedShort` model. Write unit tests for caption generation.
+
+Steps:
+1. Read `backend/pipeline/highlight_scorer.py` for `extract_word_timings` signature and output format. Read `backend/pipeline/shorts_generator.py` for `extract_clip` and `PRESETS`. Read `backend/pipeline/stages.py:2869-2990` for `stage_generate_shorts` flow.
+2. Create `backend/pipeline/caption_generator.py`:
+   - `generate_ass_captions(word_timings: list[dict], clip_start: float, style_config: dict | None = None) -> str` — returns ASS file content as string. Each word gets a `Dialogue` line. Use `{\k}` karaoke tags for word-by-word highlight timing. Style: bold white text, centered bottom 15%, black outline. Offset all word times by `-clip_start` to make them clip-relative.
+   - `write_ass_file(ass_content: str, output_path: Path) -> Path` — writes to disk, returns path.
+3. Modify `extract_clip()` in `shorts_generator.py`: add optional `ass_path: Path | None = None` parameter. When provided, append `,ass={ass_path}` to the `vf_filter` string before passing to ffmpeg. Ensure the ASS filter comes after scale/pad filters.
+4. Add `captions_enabled: Mapped[bool] = mapped_column(default=False, server_default='false')` to `GeneratedShort` in `models.py`.
+5. Create Alembic migration `027_add_captions_enabled.py` for the new column.
+6. Modify `stage_generate_shorts` in `stages.py`:
+   - After loading the highlight, load `source_video.transcript_path` and parse transcript JSON (reuse the pattern from line ~2465).
+   - Call `extract_word_timings(transcript_data, clip_start, clip_end)` to get word timings for the clip window.
+   - If word timings are non-empty, call `generate_ass_captions()` and `write_ass_file()` to a temp path.
+   - Pass the ASS path to `extract_clip()`. Set `short.captions_enabled = True`.
+   - If word timings are empty, log a warning and proceed without captions.
+7. Create `backend/pipeline/test_caption_generator.py` with tests:
+   - Valid word timings → correct ASS output with proper timing math
+   - Empty word timings → empty ASS (or raise, depending on design)
+   - Clip offset applied correctly (word at t=10.5 with clip_start=10.0 becomes t=0.5)
+   - ASS format structure (header, style block, dialogue lines)
+8. Run tests: `cd backend && python -m pytest pipeline/test_caption_generator.py -v`
+
+## Inputs
+
+- ``backend/pipeline/highlight_scorer.py` — extract_word_timings function signature and output format`
+- ``backend/pipeline/shorts_generator.py` — extract_clip function to modify`
+- ``backend/pipeline/stages.py` — stage_generate_shorts function to wire caption generation into`
+- ``backend/models.py` — GeneratedShort model to add captions_enabled column`
+
+## Expected Output
+
+- ``backend/pipeline/caption_generator.py` — new module: ASS subtitle generation from word timings`
+- ``backend/pipeline/shorts_generator.py` — modified: extract_clip accepts optional ass_path parameter`
+- ``backend/pipeline/stages.py` — modified: stage_generate_shorts loads transcript, generates captions`
+- ``backend/models.py` — modified: GeneratedShort has captions_enabled boolean`
+- ``alembic/versions/027_add_captions_enabled.py` — new migration for captions_enabled column`
+- ``backend/pipeline/test_caption_generator.py` — new: unit tests for caption generator`
+
+## Verification
+
+cd backend && python -m pytest pipeline/test_caption_generator.py -v && python -c "from pipeline.caption_generator import generate_ass_captions; print('import ok')"
diff --git a/.gsd/milestones/M024/slices/S04/tasks/T01-SUMMARY.md b/.gsd/milestones/M024/slices/S04/tasks/T01-SUMMARY.md
new file mode 100644
index 0000000..ea6174b
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S04/tasks/T01-SUMMARY.md
@@ -0,0 +1,87 @@
+---
+id: T01
+parent: S04
+milestone: M024
+provides: []
+requires: []
+affects: []
+key_files: ["backend/pipeline/caption_generator.py", "backend/pipeline/shorts_generator.py", "backend/pipeline/stages.py", "backend/models.py", "alembic/versions/027_add_captions_enabled.py", "backend/pipeline/test_caption_generator.py"]
+key_decisions: ["ASS karaoke format with per-word Dialogue lines and \k tags", "Caption generation failures are non-blocking — shorts proceed without captions", "Single ASS file shared across all format presets"]
+patterns_established: []
+drill_down_paths: []
+observability_surfaces: []
+duration: ""
+verification_result: "17 unit tests pass covering time formatting, ASS structure, clip offset math, karaoke duration calculation, empty/whitespace word handling, custom style overrides, negative time clamping, and file I/O. Import verification confirms module loads correctly."
+completed_at: 2026-04-04T11:12:15.208Z
+blocker_discovered: false
+---
+
+# T01: Created ASS subtitle generator with karaoke word-by-word highlighting and wired it into the shorts generation stage with non-blocking caption enrichment
+
+> Created ASS subtitle generator with karaoke word-by-word highlighting and wired it into the shorts generation stage with non-blocking caption enrichment
+
+## What Happened
+---
+id: T01
+parent: S04
+milestone: M024
+key_files:
+  - backend/pipeline/caption_generator.py
+  - backend/pipeline/shorts_generator.py
+  - backend/pipeline/stages.py
+  - backend/models.py
+  - alembic/versions/027_add_captions_enabled.py
+  - backend/pipeline/test_caption_generator.py
+key_decisions:
+  - ASS karaoke format with per-word Dialogue lines and \k tags
+  - Caption generation failures are non-blocking — shorts proceed without captions
+  - Single ASS file shared across all format presets
+duration: ""
+verification_result: passed
+completed_at: 2026-04-04T11:12:15.209Z
+blocker_discovered: false
+---
+
+# T01: Created ASS subtitle generator with karaoke word-by-word highlighting and wired it into the shorts generation stage with non-blocking caption enrichment
+
+**Created ASS subtitle generator with karaoke word-by-word highlighting and wired it into the shorts generation stage with non-blocking caption enrichment**
+
+## What Happened
+
+Built caption_generator.py with generate_ass_captions() and write_ass_file() for ASS subtitle generation with \k karaoke tags. Modified extract_clip() to accept optional ass_path parameter for subtitle burn-in via ffmpeg. Added captions_enabled boolean to GeneratedShort model with migration 027. Wired transcript loading and caption generation into stage_generate_shorts with non-blocking error handling — caption failures log WARNING but don't fail the stage.
+
+## Verification
+
+17 unit tests pass covering time formatting, ASS structure, clip offset math, karaoke duration calculation, empty/whitespace word handling, custom style overrides, negative time clamping, and file I/O. Import verification confirms module loads correctly.
+
+## Verification Evidence
+
+| # | Command | Exit Code | Verdict | Duration |
+|---|---------|-----------|---------|----------|
+| 1 | `cd backend && python -m pytest pipeline/test_caption_generator.py -v` | 0 | ✅ pass | 20ms |
+| 2 | `cd backend && python -c "from pipeline.caption_generator import generate_ass_captions; print('import ok')"` | 0 | ✅ pass | 100ms |
+
+
+## Deviations
+
+None.
+
+## Known Issues
+
+None.
+
+## Files Created/Modified
+
+- `backend/pipeline/caption_generator.py`
+- `backend/pipeline/shorts_generator.py`
+- `backend/pipeline/stages.py`
+- `backend/models.py`
+- `alembic/versions/027_add_captions_enabled.py`
+- `backend/pipeline/test_caption_generator.py`
+
+
+## Deviations
+None.
+
+## Known Issues
+None.
diff --git a/.gsd/milestones/M024/slices/S04/tasks/T02-PLAN.md b/.gsd/milestones/M024/slices/S04/tasks/T02-PLAN.md
new file mode 100644
index 0000000..2ae5c79
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S04/tasks/T02-PLAN.md
@@ -0,0 +1,48 @@
+---
+estimated_steps: 19
+estimated_files: 6
+skills_used: []
+---
+
+# T02: Build card renderer and concat pipeline for intro/outro templates
+
+Create `card_renderer.py` that generates intro/outro card video segments using ffmpeg lavfi (color + drawtext). Add `shorts_template` JSONB column to Creator model. Implement ffmpeg concat demuxer logic to assemble intro + main clip + outro into final short. Wire into `stage_generate_shorts`. Write unit tests for card renderer.
+
+Steps:
+1. Read T01 outputs: `backend/pipeline/caption_generator.py`, modified `shorts_generator.py` and `stages.py`.
+2. Add `shorts_template: Mapped[dict | None] = mapped_column(JSONB, nullable=True)` to `Creator` model in `models.py`. Create Alembic migration `028_add_shorts_template.py`.
+3. Create `backend/pipeline/card_renderer.py`:
+   - `render_card(text: str, duration_secs: float, width: int, height: int, accent_color: str = '#22d3ee', font_family: str = 'Inter') -> list[str]` — returns ffmpeg command args that generate a card mp4 from lavfi input (`color=c=black:s={w}x{h}:d={dur}` with `drawtext` for centered text, accent color underline/glow).
+   - `render_card_to_file(text: str, duration_secs: float, width: int, height: int, output_path: Path, accent_color: str = '#22d3ee', font_family: str = 'Inter') -> Path` — executes the ffmpeg command, returns output path.
+   - `concat_segments(segments: list[Path], output_path: Path) -> Path` — writes a concat demuxer list file, runs `ffmpeg -f concat -safe 0 -i list.txt -c copy output.mp4`, returns output path. All segments must share codec settings.
+4. Modify `shorts_generator.py`: add `extract_clip_with_template(input_path, output_path, start_secs, end_secs, vf_filter, ass_path=None, intro_path=None, outro_path=None) -> None` that extracts the main clip (with optional captions), then if intro/outro paths are provided, concats them via `concat_segments()`.
+5. Modify `stage_generate_shorts` in `stages.py`:
+   - After loading highlight, also load `highlight.source_video.creator` to access `creator.shorts_template`.
+   - If `shorts_template` exists and `show_intro` is true, call `render_card_to_file()` for intro. Same for outro.
+   - Pass intro/outro paths to the clip extraction. Use codec-compatible settings (libx264, aac, same resolution from preset spec).
+   - If no template, proceed without cards (existing behavior preserved).
+6. Create `backend/pipeline/test_card_renderer.py` with tests:
+   - `render_card()` returns valid ffmpeg command with correct dimensions and duration
+   - `concat_segments()` generates correct concat list file content
+   - Template config parsing handles missing/partial fields with defaults
+7. Run tests: `cd backend && python -m pytest pipeline/test_card_renderer.py -v`
+
+## Inputs
+
+- ``backend/pipeline/caption_generator.py` — T01 output, needed for understanding the full pipeline flow`
+- ``backend/pipeline/shorts_generator.py` — T01 output with ass_path support, to add template support`
+- ``backend/pipeline/stages.py` — T01 output with transcript loading, to add template loading`
+- ``backend/models.py` — T01 output with captions_enabled, to add shorts_template column`
+
+## Expected Output
+
+- ``backend/pipeline/card_renderer.py` — new module: ffmpeg-based intro/outro card generation + concat`
+- ``backend/pipeline/shorts_generator.py` — modified: extract_clip_with_template function`
+- ``backend/pipeline/stages.py` — modified: stage_generate_shorts loads creator template, renders cards`
+- ``backend/models.py` — modified: Creator has shorts_template JSONB column`
+- ``alembic/versions/028_add_shorts_template.py` — new migration for shorts_template column`
+- ``backend/pipeline/test_card_renderer.py` — new: unit tests for card renderer and concat`
+
+## Verification
+
+cd backend && python -m pytest pipeline/test_card_renderer.py -v && python -c "from pipeline.card_renderer import render_card, concat_segments; print('import ok')"
diff --git a/.gsd/milestones/M024/slices/S04/tasks/T03-PLAN.md b/.gsd/milestones/M024/slices/S04/tasks/T03-PLAN.md
new file mode 100644
index 0000000..900e3cf
--- /dev/null
+++ b/.gsd/milestones/M024/slices/S04/tasks/T03-PLAN.md
@@ -0,0 +1,47 @@
+---
+estimated_steps: 17
+estimated_files: 6
+skills_used: []
+---
+
+# T03: Template API endpoints and frontend template config UI
+
+Add REST endpoints for reading and updating creator shorts template config. Add template configuration UI to the HighlightQueue page — color picker, text inputs, duration controls, and intro/outro toggles. Add a caption toggle to the short generation flow.
+
+Steps:
+1. Read T02 outputs to understand the shorts_template schema on the Creator model.
+2. Create or extend `backend/routers/creators.py` with two endpoints:
+   - `GET /api/v1/admin/creators/{creator_id}/shorts-template` — returns current `shorts_template` JSONB or default config if null.
+   - `PUT /api/v1/admin/creators/{creator_id}/shorts-template` — validates and saves template config. Pydantic schema: `ShortsTemplateUpdate` with fields: intro_text (str, max 100), outro_text (str, max 100), accent_color (str, hex pattern), font_family (str), intro_duration_secs (float, 1.0-5.0), outro_duration_secs (float, 1.0-5.0), show_intro (bool), show_outro (bool).
+   - Both endpoints require admin auth.
+3. Add `ShortsTemplateConfig` and `ShortsTemplateUpdate` Pydantic schemas to `backend/schemas.py`.
+4. Create `frontend/src/api/templates.ts` — API client functions: `fetchShortsTemplate(creatorId)`, `updateShortsTemplate(creatorId, config)`.
+5. Add template config UI to `frontend/src/pages/HighlightQueue.tsx`:
+   - A collapsible "Shorts Template" section in the sidebar or above the queue.
+   - Fields: intro text, outro text, accent color (HTML color input), intro/outro duration sliders (1-5s), show intro/outro toggles.
+   - Save button that calls `updateShortsTemplate()`.
+   - Load current template on mount when a creator is selected.
+6. Add a "Captions" toggle checkbox to the short generation trigger in HighlightQueue — when unchecked, pass `captions=false` query param to the generate endpoint. Update the `POST /api/v1/admin/highlights/{id}/generate-shorts` handler (in `backend/routers/creator_highlights.py` or similar) to accept and forward an optional `captions` param.
+7. Verify frontend builds: `cd frontend && npm run build`
+8. Verify API imports: `cd backend && python -c "from routers.creators import router; print('ok')"`
+
+## Inputs
+
+- ``backend/models.py` — T02 output with shorts_template JSONB on Creator model`
+- ``backend/pipeline/card_renderer.py` — T02 output, understanding template schema`
+- ``frontend/src/pages/HighlightQueue.tsx` — existing page to add template config UI`
+- ``frontend/src/api/shorts.ts` — existing shorts API client for reference`
+- ``backend/routers/creators.py` — existing router to add template endpoints`
+
+## Expected Output
+
+- ``backend/routers/creators.py` — modified: GET/PUT shorts-template endpoints`
+- ``backend/schemas.py` — modified: ShortsTemplateConfig and ShortsTemplateUpdate schemas`
+- ``frontend/src/api/templates.ts` — new: API client for template config`
+- ``frontend/src/pages/HighlightQueue.tsx` — modified: template config UI section + caption toggle`
+- ``frontend/src/pages/HighlightQueue.module.css` — modified: styles for template config UI`
+- ``backend/routers/creator_highlights.py` — modified: captions param on generate-shorts endpoint`
+
+## Verification
+
+cd frontend && npm run build 2>&1 | tail -5 && cd ../backend && python -c "from routers.creators import router; print('ok')"
diff --git a/alembic/versions/027_add_captions_enabled.py b/alembic/versions/027_add_captions_enabled.py
new file mode 100644
index 0000000..6d71df5
--- /dev/null
+++ b/alembic/versions/027_add_captions_enabled.py
@@ -0,0 +1,30 @@
+"""Add captions_enabled boolean to generated_shorts.
+
+Revision ID: 027_add_captions_enabled
+Revises: 026_add_share_token
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "027_add_captions_enabled"
+down_revision = "026_add_share_token"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "generated_shorts",
+        sa.Column(
+            "captions_enabled",
+            sa.Boolean(),
+            nullable=False,
+            server_default=sa.text("false"),
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("generated_shorts", "captions_enabled")
diff --git a/backend/models.py b/backend/models.py
index 11e03fa..0d9d72c 100644
--- a/backend/models.py
+++ b/backend/models.py
@@ -867,5 +867,9 @@ class GeneratedShort(Base):
         default=_now, server_default=func.now(), onupdate=_now
     )
 
+    captions_enabled: Mapped[bool] = mapped_column(
+        Boolean, default=False, server_default=text("'false'"),
+    )
+
     # relationships
     highlight_candidate: Mapped[HighlightCandidate] = sa_relationship()
diff --git a/backend/pipeline/caption_generator.py b/backend/pipeline/caption_generator.py
new file mode 100644
index 0000000..2ff3b15
--- /dev/null
+++ b/backend/pipeline/caption_generator.py
@@ -0,0 +1,155 @@
+r"""ASS (Advanced SubStation Alpha) caption generator for shorts.
+
+Converts word-level timings from Whisper transcripts into ASS subtitle
+files with word-by-word karaoke highlighting. Each word gets its own
+Dialogue line with {\k} tags that control highlight duration.
+
+Pure functions — no DB access, no Celery dependency.
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+# ── Default style configuration ──────────────────────────────────────────────
+
+DEFAULT_STYLE: dict[str, Any] = {
+    "font_name": "Arial",
+    "font_size": 48,
+    "primary_colour": "&H00FFFFFF",   # white (BGR + alpha)
+    "secondary_colour": "&H0000FFFF",  # yellow highlight
+    "outline_colour": "&H00000000",   # black outline
+    "back_colour": "&H80000000",      # semi-transparent black shadow
+    "bold": -1,                       # bold
+    "outline": 3,
+    "shadow": 1,
+    "alignment": 2,                   # bottom-center
+    "margin_v": 60,                   # 60px from bottom (~15% on 1920h)
+}
+
+
+def _format_ass_time(seconds: float) -> str:
+    """Convert seconds to ASS timestamp format: H:MM:SS.cc (centiseconds).
+
+    >>> _format_ass_time(65.5)
+    '0:01:05.50'
+    >>> _format_ass_time(0.0)
+    '0:00:00.00'
+    """
+    if seconds < 0:
+        seconds = 0.0
+    h = int(seconds // 3600)
+    m = int((seconds % 3600) // 60)
+    s = seconds % 60
+    return f"{h}:{m:02d}:{s:05.2f}"
+
+
+def _build_ass_header(style_config: dict[str, Any]) -> str:
+    """Build ASS file header with script info and style definition."""
+    cfg = {**DEFAULT_STYLE, **(style_config or {})}
+
+    header = (
+        "[Script Info]\n"
+        "Title: Chrysopedia Auto-Captions\n"
+        "ScriptType: v4.00+\n"
+        "PlayResX: 1080\n"
+        "PlayResY: 1920\n"
+        "WrapStyle: 0\n"
+        "ScaledBorderAndShadow: yes\n"
+        "\n"
+        "[V4+ Styles]\n"
+        "Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, "
+        "OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, "
+        "ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, "
+        "Alignment, MarginL, MarginR, MarginV, Encoding\n"
+        f"Style: Default,{cfg['font_name']},{cfg['font_size']},"
+        f"{cfg['primary_colour']},{cfg['secondary_colour']},"
+        f"{cfg['outline_colour']},{cfg['back_colour']},"
+        f"{cfg['bold']},0,0,0,"
+        f"100,100,0,0,1,{cfg['outline']},{cfg['shadow']},"
+        f"{cfg['alignment']},20,20,{cfg['margin_v']},1\n"
+        "\n"
+        "[Events]\n"
+        "Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text\n"
+    )
+    return header
+
+
+def generate_ass_captions(
+    word_timings: list[dict[str, Any]],
+    clip_start: float,
+    style_config: dict[str, Any] | None = None,
+) -> str:
+    """Generate ASS subtitle content from word-level timings.
+
+    Each word is emitted as a separate Dialogue line with karaoke timing
+    (``{\\k<centiseconds>}``) so that words highlight one-by-one.
+
+    All word timestamps are offset by ``-clip_start`` to make them
+    clip-relative (i.e. the first frame of the clip is t=0).
+
+    Parameters
+    ----------
+    word_timings : list[dict]
+        Word-timing dicts with ``word``, ``start``, ``end`` keys.
+        ``start`` and ``end`` are absolute times in seconds.
+    clip_start : float
+        Absolute start time of the clip in seconds. Subtracted from
+        all word timestamps.
+    style_config : dict | None
+        Override style parameters (merged onto DEFAULT_STYLE).
+
+    Returns
+    -------
+    str — Full ASS file content. Empty dialogue section if no timings.
+    """
+    header = _build_ass_header(style_config)
+
+    if not word_timings:
+        logger.debug("No word timings provided — returning header-only ASS")
+        return header
+
+    lines: list[str] = [header]
+
+    for w in word_timings:
+        word_text = w.get("word", "").strip()
+        if not word_text:
+            continue
+
+        abs_start = float(w.get("start", 0.0))
+        abs_end = float(w.get("end", abs_start))
+
+        # Make clip-relative
+        rel_start = max(0.0, abs_start - clip_start)
+        rel_end = max(rel_start, abs_end - clip_start)
+
+        # Karaoke duration in centiseconds
+        k_duration = max(1, round((rel_end - rel_start) * 100))
+
+        start_ts = _format_ass_time(rel_start)
+        end_ts = _format_ass_time(rel_end)
+
+        # Dialogue line with karaoke tag
+        line = (
+            f"Dialogue: 0,{start_ts},{end_ts},Default,,0,0,0,,"
+            f"{{\\k{k_duration}}}{word_text}"
+        )
+        lines.append(line)
+
+    return "\n".join(lines) + "\n"
+
+
+def write_ass_file(ass_content: str, output_path: Path) -> Path:
+    """Write ASS content to disk.
+
+    Creates parent directories if needed. Returns the output path.
+    """
+    output_path = Path(output_path)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text(ass_content, encoding="utf-8")
+    logger.debug("Wrote ASS captions to %s (%d bytes)", output_path, len(ass_content))
+    return output_path
diff --git a/backend/pipeline/shorts_generator.py b/backend/pipeline/shorts_generator.py
index b5f63a3..0f7c711 100644
--- a/backend/pipeline/shorts_generator.py
+++ b/backend/pipeline/shorts_generator.py
@@ -72,18 +72,24 @@ def extract_clip(
     start_secs: float,
     end_secs: float,
     vf_filter: str,
+    ass_path: Path | str | None = None,
 ) -> None:
     """Extract a clip from a video file using ffmpeg.
 
     Seeks to *start_secs*, encodes until *end_secs*, and applies *vf_filter*.
     Uses ``-c:v libx264 -preset fast -crf 23`` for reasonable quality/speed.
 
+    When *ass_path* is provided, the ASS subtitle filter is appended to the
+    video filter chain so that captions are burned into the output video.
+
     Args:
         input_path: Source video file.
         output_path: Destination mp4 file (parent dir must exist).
         start_secs: Start time in seconds.
         end_secs: End time in seconds.
         vf_filter: ffmpeg ``-vf`` filter string.
+        ass_path: Optional path to an ASS subtitle file. When provided,
+            ``ass=<path>`` is appended to the filter chain.
 
     Raises:
         subprocess.CalledProcessError: If ffmpeg exits non-zero.
@@ -97,13 +103,20 @@ def extract_clip(
             f"(duration={duration}s)"
         )
 
+    # Build the video filter chain — ASS burn-in comes after scale/pad
+    effective_vf = vf_filter
+    if ass_path is not None:
+        # Escape colons and backslashes in the path for ffmpeg filter syntax
+        escaped = str(ass_path).replace("\\", "\\\\").replace(":", "\\:")
+        effective_vf = f"{vf_filter},ass={escaped}"
+
     cmd = [
         "ffmpeg",
         "-y",                          # overwrite output
         "-ss", str(start_secs),        # seek before input (fast)
         "-i", str(input_path),
         "-t", str(duration),
-        "-vf", vf_filter,
+        "-vf", effective_vf,
         "-c:v", "libx264",
         "-preset", "fast",
         "-crf", "23",
diff --git a/backend/pipeline/stages.py b/backend/pipeline/stages.py
index c3a4e28..03c50d2 100644
--- a/backend/pipeline/stages.py
+++ b/backend/pipeline/stages.py
@@ -2876,7 +2876,8 @@ def stage_generate_shorts(self, highlight_candidate_id: str) -> str:
     Returns the highlight_candidate_id on completion.
     """
     from pipeline.shorts_generator import PRESETS, extract_clip, resolve_video_path
-    from models import FormatPreset, GeneratedShort, ShortStatus
+    from pipeline.caption_generator import generate_ass_captions, write_ass_file
+    from models import FormatPreset, GeneratedShort, ShortStatus, SourceVideo
 
     start = time.monotonic()
     session = _get_sync_session()
@@ -2954,6 +2955,56 @@ def stage_generate_shorts(self, highlight_candidate_id: str) -> str:
             clip_start, clip_end,
         )
 
+        # ── Generate captions from transcript (if available) ────────────
+        ass_path: Path | None = None
+        captions_ok = False
+        try:
+            transcript_data: list | None = None
+            if source_video.transcript_path:
+                try:
+                    with open(source_video.transcript_path, "r") as fh:
+                        raw = json.load(fh)
+                    if isinstance(raw, dict):
+                        transcript_data = raw.get("segments", raw.get("results", []))
+                    elif isinstance(raw, list):
+                        transcript_data = raw
+                except (FileNotFoundError, json.JSONDecodeError, OSError) as io_exc:
+                    logger.warning(
+                        "Failed to load transcript for captions highlight=%s: %s",
+                        highlight_candidate_id, io_exc,
+                    )
+
+            if transcript_data:
+                from pipeline.highlight_scorer import extract_word_timings
+
+                word_timings = extract_word_timings(transcript_data, clip_start, clip_end)
+                if word_timings:
+                    ass_content = generate_ass_captions(word_timings, clip_start)
+                    ass_path = write_ass_file(
+                        ass_content,
+                        Path(f"/tmp/captions_{highlight_candidate_id}.ass"),
+                    )
+                    captions_ok = True
+                    logger.info(
+                        "Generated captions for highlight=%s (%d words)",
+                        highlight_candidate_id, len(word_timings),
+                    )
+                else:
+                    logger.warning(
+                        "No word timings in transcript window [%.1f–%.1f]s for highlight=%s — proceeding without captions",
+                        clip_start, clip_end, highlight_candidate_id,
+                    )
+            else:
+                logger.info(
+                    "No transcript available for highlight=%s — proceeding without captions",
+                    highlight_candidate_id,
+                )
+        except Exception as cap_exc:
+            logger.warning(
+                "Caption generation failed for highlight=%s: %s — proceeding without captions",
+                highlight_candidate_id, cap_exc,
+            )
+
         # ── Process each preset independently ───────────────────────────
         for preset in FormatPreset:
             spec = PRESETS[preset]
@@ -2983,6 +3034,7 @@ def stage_generate_shorts(self, highlight_candidate_id: str) -> str:
                     start_secs=clip_start,
                     end_secs=clip_end,
                     vf_filter=spec.vf_filter,
+                    ass_path=ass_path,
                 )
 
                 # Upload to MinIO
@@ -3000,6 +3052,7 @@ def stage_generate_shorts(self, highlight_candidate_id: str) -> str:
                 short.status = ShortStatus.complete
                 short.file_size_bytes = file_size
                 short.minio_object_key = minio_key
+                short.captions_enabled = captions_ok
                 short.share_token = secrets.token_urlsafe(8)
                 session.commit()
 
@@ -3035,6 +3088,13 @@ def stage_generate_shorts(self, highlight_candidate_id: str) -> str:
                     except OSError:
                         pass
 
+        # Clean up temp ASS caption file
+        if ass_path is not None and ass_path.exists():
+            try:
+                ass_path.unlink()
+            except OSError:
+                pass
+
         elapsed = time.monotonic() - start
         logger.info(
             "Shorts generation complete for highlight=%s in %.1fs",
diff --git a/backend/pipeline/test_caption_generator.py b/backend/pipeline/test_caption_generator.py
new file mode 100644
index 0000000..249faac
--- /dev/null
+++ b/backend/pipeline/test_caption_generator.py
@@ -0,0 +1,159 @@
+"""Unit tests for caption_generator module."""
+
+from __future__ import annotations
+
+import re
+import tempfile
+from pathlib import Path
+
+import pytest
+
+from pipeline.caption_generator import (
+    DEFAULT_STYLE,
+    _format_ass_time,
+    generate_ass_captions,
+    write_ass_file,
+)
+
+
+# ── Fixtures ─────────────────────────────────────────────────────────────────
+
+@pytest.fixture
+def sample_word_timings() -> list[dict]:
+    """Realistic word timings as produced by extract_word_timings."""
+    return [
+        {"word": "This", "start": 10.0, "end": 10.3},
+        {"word": "is", "start": 10.3, "end": 10.5},
+        {"word": "a", "start": 10.5, "end": 10.6},
+        {"word": "test", "start": 10.6, "end": 11.0},
+        {"word": "sentence", "start": 11.1, "end": 11.6},
+    ]
+
+
+# ── Time formatting ─────────────────────────────────────────────────────────
+
+class TestFormatAssTime:
+    def test_zero(self):
+        assert _format_ass_time(0.0) == "0:00:00.00"
+
+    def test_sub_second(self):
+        assert _format_ass_time(0.5) == "0:00:00.50"
+
+    def test_minutes(self):
+        assert _format_ass_time(65.5) == "0:01:05.50"
+
+    def test_hours(self):
+        assert _format_ass_time(3661.25) == "1:01:01.25"
+
+    def test_negative_clamps_to_zero(self):
+        assert _format_ass_time(-5.0) == "0:00:00.00"
+
+
+# ── ASS generation ──────────────────────────────────────────────────────────
+
+class TestGenerateAssCaptions:
+    def test_empty_timings_returns_header_only(self):
+        result = generate_ass_captions([], clip_start=0.0)
+        assert "[Script Info]" in result
+        assert "[Events]" in result
+        # No Dialogue lines
+        assert "Dialogue:" not in result
+
+    def test_structure_has_required_sections(self, sample_word_timings):
+        result = generate_ass_captions(sample_word_timings, clip_start=10.0)
+        assert "[Script Info]" in result
+        assert "[V4+ Styles]" in result
+        assert "[Events]" in result
+        assert "Dialogue:" in result
+
+    def test_clip_offset_applied(self, sample_word_timings):
+        """Word at t=10.5 with clip_start=10.0 should become t=0.5 in ASS."""
+        result = generate_ass_captions(sample_word_timings, clip_start=10.0)
+        lines = result.strip().split("\n")
+        dialogue_lines = [l for l in lines if l.startswith("Dialogue:")]
+
+        # First word "This" starts at 10.0, clip_start=10.0 → relative 0.0
+        assert dialogue_lines[0].startswith("Dialogue: 0,0:00:00.00,")
+
+        # Third word "a" starts at 10.5, clip_start=10.0 → relative 0.5
+        assert "0:00:00.50" in dialogue_lines[2]
+
+    def test_karaoke_tags_present(self, sample_word_timings):
+        result = generate_ass_captions(sample_word_timings, clip_start=10.0)
+        lines = result.strip().split("\n")
+        dialogue_lines = [l for l in lines if l.startswith("Dialogue:")]
+
+        for line in dialogue_lines:
+            # Each line should have a \kN tag
+            assert re.search(r"\{\\k\d+\}", line), f"Missing karaoke tag in: {line}"
+
+    def test_karaoke_duration_math(self, sample_word_timings):
+        """Word "This" at [10.0, 10.3] → 0.3s → k30 (30 centiseconds)."""
+        result = generate_ass_captions(sample_word_timings, clip_start=10.0)
+        lines = result.strip().split("\n")
+        dialogue_lines = [l for l in lines if l.startswith("Dialogue:")]
+
+        # "This" duration: 10.3 - 10.0 = 0.3s = 30cs
+        assert "{\\k30}This" in dialogue_lines[0]
+
+        # "test" duration: 11.0 - 10.6 = 0.4s = 40cs
+        assert "{\\k40}test" in dialogue_lines[3]
+
+    def test_word_count_matches(self, sample_word_timings):
+        result = generate_ass_captions(sample_word_timings, clip_start=10.0)
+        lines = result.strip().split("\n")
+        dialogue_lines = [l for l in lines if l.startswith("Dialogue:")]
+        assert len(dialogue_lines) == 5
+
+    def test_empty_word_text_skipped(self):
+        timings = [
+            {"word": "hello", "start": 0.0, "end": 0.5},
+            {"word": "  ", "start": 0.5, "end": 0.7},  # whitespace-only
+            {"word": "", "start": 0.7, "end": 0.8},     # empty
+            {"word": "world", "start": 0.8, "end": 1.2},
+        ]
+        result = generate_ass_captions(timings, clip_start=0.0)
+        lines = result.strip().split("\n")
+        dialogue_lines = [l for l in lines if l.startswith("Dialogue:")]
+        assert len(dialogue_lines) == 2  # only "hello" and "world"
+
+    def test_custom_style_overrides(self, sample_word_timings):
+        result = generate_ass_captions(
+            sample_word_timings,
+            clip_start=10.0,
+            style_config={"font_size": 72, "font_name": "Roboto"},
+        )
+        assert "Roboto" in result
+        assert ",72," in result
+
+    def test_negative_relative_time_clamped(self):
+        """Words before clip_start should clamp to 0."""
+        timings = [{"word": "early", "start": 5.0, "end": 5.5}]
+        result = generate_ass_captions(timings, clip_start=10.0)
+        lines = [l for l in result.strip().split("\n") if l.startswith("Dialogue:")]
+        # Both start and end clamped to 0
+        assert lines[0].startswith("Dialogue: 0,0:00:00.00,0:00:00.00,")
+
+
+# ── File writing ─────────────────────────────────────────────────────────────
+
+class TestWriteAssFile:
+    def test_writes_content(self):
+        content = "[Script Info]\ntest content\n"
+        with tempfile.TemporaryDirectory() as td:
+            out = write_ass_file(content, Path(td) / "sub.ass")
+            assert out.exists()
+            assert out.read_text() == content
+
+    def test_creates_parent_dirs(self):
+        content = "test"
+        with tempfile.TemporaryDirectory() as td:
+            out = write_ass_file(content, Path(td) / "nested" / "deep" / "sub.ass")
+            assert out.exists()
+
+    def test_returns_path(self):
+        content = "test"
+        with tempfile.TemporaryDirectory() as td:
+            target = Path(td) / "sub.ass"
+            result = write_ass_file(content, target)
+            assert result == target