moizibnyousaf /tiktok-slideshow — End-to-End TikTok Content Pipeline
Orchestrator that chains 3 atomic skills into a complete TikTok production pipeline:
/slideshow-script → /paper-marketing → /video-content
(5 scripts) (5 designs) (5 videos)
Each skill is a Lego block that works independently. This orchestrator is one recipe.
Architecture
Phase 1: SCRIPT — /slideshow-script generates 5 narrative scripts
Phase 2: DESIGN — /paper-marketing creates 5 visual designs (1 per script)
Phase 3: VIDEO — /video-content assembles videos from designs
Platform spec: TikTok 1080×1920, 9:16, 30fps, H.264
Prerequisites
Before starting, verify all 3 sub-skills are installed:
~/.claude/skills/slideshow-script/SKILL.md~/.claude/skills/paper-marketing/SKILL.md~/.claude/skills/video-content/SKILL.md
If any sub-skill is missing, tell the user and recommend mktg update to install it. Do not proceed without all 3.
Workflow
Phase 1: SCRIPT
Load the /slideshow-script skill at ~/.claude/skills/slideshow-script/SKILL.md and follow its workflow:
- Read brand files (positioning, audience, voice)
- User selects positioning angle
- Generate up to 5 scripts using 5 storytelling frameworks (AIDA, PAS, BAB, Star-Story-Solution, Stat-Flip). User may select fewer at the gate.
- Each script uses the brand voice from
brand/voice-profile.mdfor tone, vocabulary, and signature phrases - Present all 5 scripts for approval
- Write 5 content spec YAMLs to
marketing/content-specs/
Gate: User must approve scripts before Phase 2.
Phase 2: DESIGN
Load the /paper-marketing skill at ~/.claude/skills/paper-marketing/SKILL.md and follow its workflow, with these orchestrator-specific instructions:
- Paper-marketing Phase 1 (Load Brand) proceeds as normal
- Skip Phase 2a-2d — the content specs already exist from Phase 1
- Instead, read the 5 content spec YAMLs from
marketing/content-specs/ - Each content spec has a
visual_directionfield — use that to set each agent's design brief - Each agent gets its content spec's unique script (different slides, different narrative)
- Each agent loads the
/frontend-designskill for design quality - Spawn agents — number depends on how many scripts the user approved (could be 3-5)
Script-to-design mapping from content specs:
| Content Spec | Visual Direction | Framework |
|---|---|---|
{project}-aida.yaml |
typographic | AIDA |
{project}-pas.yaml |
contrast-play | PAS |
{project}-bab.yaml |
atmospheric | BAB |
{project}-story.yaml |
editorial | Star-Story-Solution |
{project}-statflip.yaml |
data-led | Stat-Flip |
- User reviews all designs, selects favorite(s)
- Extract
get_jsx()from selected artboard(s) - Write handoff YAML(s) to
marketing/handoffs/
Gate: User must select variation(s) and export PNG from Paper UI.
Phase 3: VIDEO
Load the /video-content skill at ~/.claude/skills/video-content/SKILL.md and follow its workflow:
- Detect handoff YAML from Phase 2
- ffmpeg slice the exported PNG into individual slides
- User selects tier (v1 Quick / v1.5 Enhanced / v2 Full)
- Assemble video at chosen tier
- ffmpeg post-process (thumbnail, GIF preview, platform encode)
For multiple selections: If user approved multiple designs, run video assembly for each. Can be parallelized — each video is independent.
Gate: User approves final video(s).
Human-in-the-Loop Gates
[Phase 1] → User approves 5 scripts
[Phase 2] → User selects design variation(s), exports PNG(s) from Paper UI
[Phase 3] → User selects video tier, approves final output
Each gate is an AskUserQuestion. The user can go back to any phase.
Output
marketing/content-specs/
{project}-aida.yaml
{project}-pas.yaml
{project}-bab.yaml
{project}-story.yaml
{project}-statflip.yaml
marketing/handoffs/
{project}-{variation}-handoff.yaml
marketing/video/
{project}-{variation}/
slides/
output_{tier}.mp4
thumbnail.png
preview.gif
Reusable Blocks
Each skill in this chain works independently:
| Block | Standalone Use |
|---|---|
/slideshow-script |
Generate scripts for any format (Instagram carousel, YouTube, email) |
/paper-marketing |
Design any visual content (not just slideshows) |
/video-content |
Assemble video from any PNGs (not just Paper exports) |
Other Orchestrator Recipes (Future)
Same blocks, different combinations:
/instagram-carousel = /slideshow-script → /paper-marketing (4:5 ratio, no video)
/youtube-short = /slideshow-script → /paper-marketing → /video-content (16:9)
/reels-batch = /slideshow-script (10 scripts) → /paper-marketing → /video-content × 10
/ad-creative = /slideshow-script (1 script) → /paper-marketing (1 slide) → /video-content
Error Recovery
| Failure | Action |
|---|---|
| Sub-skill not installed | Stop, recommend mktg update, do not proceed |
| Phase 1 scripts rejected | Revise scripts with user feedback, do not advance to Phase 2 |
| Paper MCP unavailable | Fall back to manual slide creation; user provides PNGs |
| Phase 2 design export fails | User re-exports from Paper UI; agent retries video assembly |
| Phase 3 video assembly fails | Check ffmpeg/Remotion installation, retry with v1 tier first |
Anti-Patterns
| Anti-Pattern | Instead |
|---|---|
| Skipping human gates between phases | Always wait for user approval at each gate |
| Running all 3 phases without stopping | Each phase has an explicit gate — never auto-advance |
| Generating all 5 when user only wants 2 | Respect user selection at Phase 1 gate |
| Re-using the same visual direction for all designs | Each content spec maps to a unique visual direction |
| Loading sub-skills via Bash instead of reading SKILL.md | Load each skill's SKILL.md and follow its instructions |
Principles
- Skills never call skills — this SKILL.md teaches the agent the sequence; the agent loads each skill
- Filesystem is the bus — content specs and handoffs are YAML files, not API calls
- Human gates at every phase — no runaway agent chains
- Progressive quality — start with v1 to verify, upgrade to v2 for production
- All outputs are publishable — different scripts, different designs, not N versions of 1