diff --git a/content-creation/SKILL.md b/content-creation/SKILL.md
new file mode 100644
index 0000000..cc9c779
--- /dev/null
+++ b/content-creation/SKILL.md
@@ -0,0 +1,234 @@
+---
+name: content-creation
+description: Create social videos, TikToks, Reels, and visual content using AI-generated images, video, captions, and audio. Use this skill whenever the user asks to create content, make a video, generate an image, produce a TikTok or Reel, create a promotional clip, add captions to a video, make visual content, or create short-form video. Also use when the user wants to iterate on content quality — regenerate images, try different audio, adjust text, or upscale for higher quality. Templates are optional — pass your own prompts for full creative control, or use templates as curated shortcuts.
+---
+
+# Content Creation
+
+Create social-ready short-form videos by composing independent primitives. Each primitive does one thing. You orchestrate them.
+
+## Primitives
+
+Every primitive works without a template. Pass your own prompt, reference images, and style rules directly. Templates are optional shortcuts — curated creative recipes that pre-fill parameters for a specific look.
+
+Seven primitives, each callable independently:
+
+| Primitive | Command | Async? |
+|-----------|---------|--------|
+| Transcribe audio | `recoup content transcribe` | Yes (30-60s) |
+| Generate image | `recoup content image` | Yes (10-30s) |
+| Generate video | `recoup content video` | Yes (60-180s) |
+| Generate caption | `recoup content caption` | No (2-5s) |
+| Edit | `recoup content edit` | Yes (10-30s) |
+| Upscale | `recoup content upscale` | Yes (30-60s) |
+| Analyze video | `recoup content analyze` | No (10-30s) |
+
+Async primitives return a `runId`. Poll with `recoup tasks status --run {runId} --json` until `status` is `COMPLETED`, then read `output`.
+
+There is also `recoup content create` which runs the full pipeline in one shot — use it when the user just wants a video without creative control.
+
+## Video Generation Modes
+
+`generate-video` supports 6 modes. Set `--mode` to be explicit, or omit it and the mode is inferred from the inputs you provide.
+
+| Mode | What it does | Required inputs |
+|------|-------------|-----------------|
+| `prompt` | Create a video from a text description | `--prompt` |
+| `animate` | Animate a still image | `--image`, `--prompt` |
+| `reference` | Use an image as a style/subject reference (not the first frame) | `--image`, `--prompt` |
+| `extend` | Continue an existing video (max 8s input) | `--video`, `--prompt` |
+| `first-last` | Generate a video that transitions between two images | `--image`, `--end-image`, `--prompt` |
+| `lipsync` | Sync face movement to an audio clip | `--image`, `--audio` |
+
+All modes support `--aspect-ratio` (auto, 16:9, 9:16), `--duration` (4s-8s), and `--resolution` (720p, 1080p, 4k).
+
+### When to use which mode
+
+- "I need a cinematic establishing shot" → `prompt`
+- "I have a photo of the artist, make it come alive" → `animate`
+- "I have a photo of the artist, put them in a new scene" → `reference`
+- "This clip is great but too short" → `extend`
+- "I want a smooth transition from the studio to the stage" → `first-last`
+- "I need the face to sing along to this audio" → `lipsync`
+
+## Workflow
+
+### Without a template (full control)
+
+```bash
+# 1. Transcribe audio
+recoup content transcribe --url {audioUrl} --json
+# Returns: audioUrl, fullLyrics, segments (timestamped words)
+
+# 2. Generate image
+recoup content image --prompt "{scene description}" \
+  --reference-image {referenceImageUrl} --json
+# Returns: imageUrl, images (array if num_images > 1)
+
+# 3. (Optional) Upscale image
+recoup content upscale --url {imageUrl} --type image --json
+
+# 4. Generate video
+recoup content video --mode animate --image {imageUrl} \
+  --prompt "{how to animate}" --json
+# Or from scratch:
+recoup content video --mode prompt --prompt "{scene}" --json
+# Returns: videoUrl, mode
+
+# 5. Generate caption
+recoup content caption --topic "{topic}" --length short --json
+# Synchronous — returns { content } immediately
+
+# 6. Edit final video
+recoup content edit --url {videoUrl} --audio {audioUrl} \
+  --template artist-caption-bedroom \
+  --overlay-text "{caption}" --json
+
+# 7. (Optional) Analyze the result
+recoup content analyze --video {videoUrl} --prompt "Rate this content 1-10 for social media engagement. Note any quality issues." --json
+```
+
+### With a template (shortcuts)
+
+Templates pre-fill prompts, reference images, and style rules. You can override any parameter.
+
+```bash
+recoup content create --artist {artistId} --template artist-caption-bedroom --json
+```
+
+### Lipsync
+
+```bash
+recoup content video --mode lipsync \
+  --image {faceUrl} --audio {audioUrl} --json
+
+# Edit — no need to mux audio (already baked into the video)
+recoup content edit --url {videoUrl} \
+  --crop-aspect 9:16 \
+  --overlay-text "{caption}" --json
+```
+
+### Extend a clip
+
+```bash
+recoup content video --mode extend \
+  --video {shortClipUrl} --prompt "continue the scene naturally" --json
+```
+
+### Transition between two shots
+
+```bash
+recoup content video --mode first-last \
+  --image {startFrameUrl} --end-image {endFrameUrl} \
+  --prompt "smooth cinematic transition" --json
+```
+
+## Model Selection
+
+Each generative primitive accepts an optional `--model` flag to override the default:
+
+| Primitive | Default Model | Flag |
+|-----------|---------------|------|
+| generate-image (prompt only) | `fal-ai/nano-banana-2` | `--model {modelId}` |
+| generate-image (with reference) | `fal-ai/nano-banana-2/edit` | `--model {modelId}` |
+| generate-video (prompt) | `fal-ai/veo3.1/text-to-video` | `--model {modelId}` |
+| generate-video (animate) | `fal-ai/veo3.1/image-to-video` | `--model {modelId}` |
+| generate-video (reference) | `fal-ai/veo3.1/reference-to-video` | `--model {modelId}` |
+| generate-video (extend) | `fal-ai/veo3.1/extend-video` | `--model {modelId}` |
+| generate-video (first-last) | `fal-ai/veo3.1/first-last-frame-to-video` | `--model {modelId}` |
+| generate-video (lipsync) | `fal-ai/ltx-2-19b/audio-to-video` | `--model {modelId}` |
+| transcribe-audio | `fal-ai/whisper` | `--model {modelId}` |
+
+## Edit Operations
+
+The `edit` command accepts either a template (deterministic config) or manual flags. Point `--url` at image or video — both are supported.
+
+| Operation | Flags | What it does |
+|-----------|-------|-------------|
+| Trim | `--trim-start {s} --trim-duration {s}` | Cut a time window |
+| Crop | `--crop-aspect {ratio}` | Crop to aspect ratio (e.g. 9:16) |
+| Overlay text | `--overlay-text {text} --text-color {color} --text-position {pos}` | Add on-screen text |
+| Mux audio | `--mux-audio {url}` | Add audio track to video |
+| Template | `--template {name}` | Use template's preset operations |
+
+All operations run in a single processing pass.
+
+## Template Override Priority
+
+When using a template, your explicit params always win:
+
+1. **Your flags** — highest priority. `--prompt`, `--reference-image`, etc. override everything.
+2. **Artist context** — if the artist has a style guide, it personalizes the template.
+3. **Template defaults** — lowest priority. The recipe's built-in values.
+
+Example: Use the bedroom template but override the prompt:
+
+```bash
+recoup content image --template artist-caption-bedroom \
+  --prompt "A candid selfie in a sunlit kitchen, warm tones" --json
+```
+
+The template provides reference images and style rules, but your prompt replaces the template's default prompt.
+
+## Watching Your Work
+
+You can't see pixels — but `analyze-video` can. Use it freely throughout the workflow to evaluate your output. Write your own prompts based on what you need to know.
+
+### What to look for
+
+**QA** — technical quality problems:
+- Visual artifacts, glitches, distortion
+- Subject consistency and recognizability
+- Motion naturalness vs robotic movement
+- Broken or repeated frames
+
+**Taste** — creative quality:
+- Does the opening hook attention in the first 2 seconds?
+- Does it feel intentional and aesthetic, or generic?
+- Does the visual energy match the audio energy?
+- Would a real person post this?
+
+**Platform readiness** — practical details:
+- Text readability and positioning (not blocking the subject, not cut off by UI)
+- Vertical crop framing
+- Audio-visual sync and pacing
+
+### When to analyze
+
+Use it at every creative checkpoint, not just once at the end:
+
+- **After generating a clip** — catch quality issues before spending time editing
+- **After each edit pass** — did the trim land on the right beat? Does the cut feel natural?
+- **After combining multiple clips** — do the cuts flow? Does the pacing hold attention?
+- **After adding audio** — is the lipsync convincing? Does the music energy match the visuals?
+- **After adding text** — is it readable, well-timed, not blocking anything important?
+- **When comparing versions** — which of two outputs is better and why?
+- **When building a storyline** — does the sequence of shots tell a coherent story?
+
+For longer edits with multiple cuts, analyze after each major assembly — not just the final export. If cut 3 of 5 looks wrong, fix it before adding cuts 4 and 5.
+
+### Acting on feedback
+
+- Artifacts or glitches — regenerate with a different `--model` or prompt
+- Subject not recognizable — regenerate image with a better `--reference-image` or prompt
+- Robotic motion — try a more natural prompt, switch mode, or use lipsync
+- Boring opening — more dramatic image or use `first-last` for a striking transition
+- Looks like AI slop — different model, better reference image, or more specific prompt
+- Clip too short — use `extend` mode to continue it
+- Bad text placement — adjust position or shorten caption in `edit`
+
+Generate, watch, fix, watch again. Two rounds usually gets you from mediocre to good.
+
+## Iteration
+
+Redo any step without rerunning everything:
+
+- Bad image? Run `generate-image` again with a different prompt
+- Wrong audio moment? Transcribe a different file or URL
+- Caption too long? Run `generate-caption` with `--length short`
+- Low quality? Run `upscale` on the image or video
+- Video glitchy? Check `analyze-video` feedback, then regenerate with adjusted params
+- Clip too short? Run `generate-video --mode extend` to continue it
+- Everything good but caption is wrong? Just rerun `edit` with new `--overlay-text`
+
+Each retry costs only that step's time, not the full pipeline.