@muskanmeet/invoicing-design-system 0.1.0

package/.local/secondary_skills/podcast-marketing/SKILL.md

@@ -0,0 +1,550 @@
+---
+name: podcast-marketing
+description: Repurpose podcast episodes into social posts, show notes, clips, and quote cards.
+---
+
+# Podcast Marketing
+
+Turn podcast episodes into a full suite of marketing content. Supports four input modes, extracts content atoms with viral scoring, and produces platform-ready content pieces that can be published immediately.
+
+## Requirements & Setup
+
+### Tier 1: Text Content (always needed)
+
+These are required for transcript extraction and content generation (show notes, blog, social, newsletter, pull quotes):
+
+| Dependency | Install | Purpose |
+
+|-----------|---------|---------|
+
+| Python 3.10+ | `installProgrammingLanguage({ language: "python-3.11" })` | Runtime for all scripts |
+
+| youtube-transcript-api | `installLanguagePackages({ language: "python", packages: ["youtube-transcript-api"] })` | YouTube caption extraction |
+
+| yt-dlp | `installLanguagePackages({ language: "python", packages: ["yt-dlp"] })` | YouTube metadata + video/audio download |
+
+| LLM API | Set up via Anthropic or OpenAI AI integration (Replit proxy — no user API key needed) | Content atom extraction and content generation |
+
+### Tier 2: Quote Card Images
+
+These are required to generate visual quote card PNG files:
+
+| Dependency | Install | Purpose |
+
+|-----------|---------|---------|
+
+| Pillow | `installLanguagePackages({ language: "python", packages: ["Pillow"] })` | Image generation |
+
+| System fonts | Already available on Nix (DejaVu Sans at `/usr/share/fonts/truetype/dejavu/`) | Typography |
+
+No API keys needed. Run `scripts/generate_quote_cards.py` directly.
+
+### Tier 3: Short-Form Video Clips (basic)
+
+These produce trimmed, vertical-reformatted video clips without animated captions:
+
+| Dependency | Install | Purpose |
+
+|-----------|---------|---------|
+
+| ffmpeg | `installSystemDependencies({ packages: ["ffmpeg"] })` | Video trimming, reformatting, subtitle burning |
+
+| yt-dlp | (already in Tier 1) | Download source video from YouTube |
+
+### Tier 4: Animated Captions (Opus Clips style)
+
+These add word-by-word highlighted captions to video clips:
+
+| Dependency | Install | Purpose |
+
+|-----------|---------|---------|
+
+| OpenAI API | Set up OpenAI AI integration OR user provides `OPENAI_API_KEY` | Whisper word-level timestamps |
+
+| Pillow | (already in Tier 2) | Caption frame generation (alternative to ASS subtitles) |
+
+The animated caption pipeline re-transcribes each trimmed clip through Whisper with `timestamp_granularities=["word"]` to get per-word timing. Without this, clips still work but use segment-level static captions instead.
+
+### Setup Checklist
+
+Before running the pipeline, verify:
+
+```python
+
+# Tier 1
+
+import youtube_transcript_api \# YouTube captions
+
+import yt_dlp \# YouTube metadata/download
+
+import anthropic \# or openai — for content generation
+
+# Tier 2 (if generating quote cards)
+
+from PIL import Image, ImageFont
+
+ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 48)
+
+# Tier 3 (if generating video clips)
+
+import subprocess
+
+subprocess.run(["ffmpeg", "-version"], capture_output=True, check=True)
+
+# Tier 4 (if generating animated captions)
+
+import openai \# OPENAI_API_KEY must be set
+
+```
+
+## When to Use
+
+- User wants to repurpose a podcast episode into marketing content
+- User asks to create social media posts, blog articles, or newsletters from an episode
+
+- User wants show notes with timestamps and key takeaways
+- User asks to generate pull quotes, audiogram clips, or short-form video clips
+
+- User provides an RSS feed and wants to process episodes in batch
+- User shares a YouTube link to a podcast episode or video interview
+
+- User says "turn this episode into content", "repurpose this podcast", or "create marketing from my episode"
+- User wants to create a content calendar from podcast episodes
+
+## When NOT to Use
+
+- Creating a podcast from scratch (use the `podcast-generator` skill instead)
+- General social media management unrelated to podcast content
+
+## Input Modes
+
+### Mode 1: RSS Feed
+
+Auto-download and transcribe the latest episode(s) from a podcast RSS feed.
+
+1. Parse the RSS feed to extract episode metadata (title, description, publish date, audio URL).
+2. Download the audio file.
+
+3. Transcribe using OpenAI Whisper (produces timestamped transcript).
+4. Proceed to content generation.
+
+Read `rss-and-batch.md` for RSS parsing details and the Whisper transcription pipeline.
+
+### Mode 2: Raw Transcript
+
+Accept a transcript file directly — plain text, SRT, or VTT format.
+
+1. Detect format from file extension or content structure.
+2. Parse into a normalized transcript (list of segments with optional timestamps).
+
+3. Proceed to content generation.
+
+SRT/VTT files preserve timestamps, which enables show notes with time markers. Plain text works but won't have timestamp data.
+
+### Mode 3: YouTube Link
+
+Extract transcript and metadata from a YouTube video — works for podcast episodes, interviews, talks, or any video with speech.
+
+1. Accept a YouTube URL (full or short format).
+2. Attempt to fetch YouTube's auto-generated or manually uploaded captions first (fast, free, no API key needed).
+
+3. If no captions are available, download the audio track and transcribe via Whisper.
+4. Extract video metadata (title, channel, description, publish date, duration).
+
+5. Proceed to content generation.
+
+Read `youtube-processing.md` for the full YouTube extraction pipeline.
+
+### Mode 4: Batch Mode
+
+Process the last N episodes from an RSS feed automatically.
+
+1. Parse the RSS feed.
+2. Sort episodes by publish date (newest first).
+
+3. Download and transcribe each episode.
+4. Run the content generation pipeline for each episode.
+
+5. Organize output by episode (one directory per episode).
+
+Read `rss-and-batch.md` for batch processing details.
+
+## Content Generation Pipeline
+
+Once you have a transcript and episode metadata, follow this pipeline. The order matters — each step builds on the previous one.
+
+### Step 1: Extract Content Atoms
+
+This replaces a simple summary. Analyze the full transcript and extract **7 types of content atoms** — these are the raw material for everything downstream.
+
+| Atom Type | What to Look For | Example |
+
+|-----------|-----------------|---------|
+
+| **Narrative Arcs** | Stories with setup, tension, resolution | "When we almost lost the company in 2019..." |
+
+| **Quotable Moments** | Punchy, memorable one-liners | "Culture eats strategy for breakfast, but execution eats both" |
+
+| **Controversial Takes** | Opinions that challenge mainstream thinking | "I think remote work is actually worse for junior employees" |
+
+| **Data Points** | Statistics, numbers, research findings | "We saw a 340% increase after switching to..." |
+
+| **Frameworks** | Mental models, step-by-step processes | "The 3-bucket prioritization system..." |
+
+| **Stories** | Personal anecdotes, case studies, examples | "Our first customer was actually..." |
+
+| **Predictions** | Forward-looking claims about trends | "In 5 years, every company will need to..." |
+
+For each atom, capture: the verbatim text, timestamps (start/end), speaker, atom type, and a one-line hook that could open a social post.
+
+**Timestamp requirement:** Every atom MUST include `start_timestamp`and`end_timestamp` in seconds. Match the verbatim text against the transcript segments to find the correct timestamps. Without timestamps, video clip selection is impossible.
+
+**Long transcript handling:** If the transcript exceeds 50,000 characters, split into overlapping chunks (5,000 char overlap) and extract atoms from each chunk separately. Deduplicate by comparing verbatim_text similarity (>80% overlap = duplicate). Merge and re-rank.
+
+Read `content-atoms.md` for the full extraction prompt and output format.
+
+### Step 2: Score and Assign Content Atoms
+
+Score each atom on three dimensions to prioritize the strongest content:
+
+| Dimension | Weight | What It Measures |
+
+|-----------|--------|-----------------|
+
+| **Novelty** | 40% | Is this new, surprising, or counter-intuitive? |
+
+| **Controversy** | 30% | Will people disagree, debate, or have strong reactions? |
+
+| **Utility** | 30% | Can someone apply this immediately? |
+
+#### Score thresholds
+
+- **80-100**: Priority publish — lead with these
+- **60-79**: Strong fill — solid content for the calendar
+
+- **40-59**: Gap filler — use when scheduling needs variety
+- **Below 40**: Cut — not worth publishing
+
+**Atom-to-platform assignment (anti-repetition rule):** After scoring, assign each atom to AT MOST 2 content pieces. No single quote should appear verbatim in more than 2 outputs. Use this mapping to assign the highest-scoring atoms first:
+
+| Platform | Best Atom Types | Lead Atom |
+
+|----------|----------------|-----------|
+
+| Twitter thread | Controversial takes, data points | Assign atom \#1 by score |
+
+| LinkedIn post | Frameworks, data points | Assign a DIFFERENT atom than Twitter |
+
+| Blog post | Narrative arcs, stories | Assign a DIFFERENT atom than Twitter/LinkedIn |
+
+| Newsletter | Top 2-3 by score, any type | Can reuse 1 atom from blog |
+
+| Quote cards | Quotable moments | Use atoms NOT already assigned to other platforms |
+
+| Video clips | Controversial takes, stories, data points | Based on timestamp availability |
+
+This ensures each platform leads with a unique angle rather than repeating the same quote everywhere.
+
+### Step 3: Show Notes
+
+Generate structured show notes with:
+
+- Episode title and one-paragraph description
+- Timestamped topic outline (if timestamps available)
+
+- Key takeaways as bullet points (5-7)
+- **ALL links mentioned** in the episode — parse from YouTube description, transcript mentions of URLs, and episode metadata. Include every reference link, not just a subset
+
+- Guest bio (if interview format — pull from transcript intro or episode metadata)
+
+### Step 4: Blog Post
+
+Write an 800-1200 word article that:
+
+- Opens with the atom assigned to the blog (from Step 2) as the hook (not "In this episode...")
+- Expands on 2-3 top atoms with context and examples from the conversation
+
+- Includes 1-2 pull quotes from the transcript
+- Ends with a CTA to listen to the full episode
+
+- Uses subheadings for scannability
+
+**Hard limit: 1,200 words maximum.** Count words before saving. If over, cut the weakest section.
+
+### Step 5: Social Media Posts
+
+Generate platform-specific posts. Each platform MUST lead with a DIFFERENT atom (from Step 2 assignment):
+
+#### Twitter/X Thread (5-8 tweets)
+
+- Tweet 1: Hook — the atom assigned to Twitter (controversial take or novel insight)
+- Tweets 2-6: One atom per tweet, using the speakers' own words when possible
+
+- Final tweet: CTA with episode link
+- **HARD LIMIT: Every tweet MUST be under 280 characters including spaces and URLs.** Count characters before saving. Split long quotes across two tweets if needed.
+
+##### LinkedIn Post (1 long-form)
+
+- Open with the atom assigned to LinkedIn (framework or data point — NOT the same as Twitter hook)
+- Share 2-3 insights with brief context
+
+- Professional tone, first-person perspective
+- End with a question to drive engagement + episode link
+
+- 1,200-1,500 characters
+
+###### Instagram Caption
+
+- Conversational, accessible tone
+- Lead with the "why should I care" angle
+
+- Include 3-5 relevant hashtags
+- End with CTA to listen (link in bio)
+
+- 300-500 characters
+
+**Multi-topic episodes:** If the episode covers 3+ distinct topics, also generate 1 standalone tweet per topic (not threads — single tweets) for drip-posting throughout the week. Include these in the calendar.
+
+Read `content-templates.md` for detailed templates and examples for each platform.
+
+### Step 6: Email Newsletter
+
+Write newsletter copy that:
+
+- Subject line: curiosity-driven, under 50 characters
+- Preview text: complements the subject line, under 90 characters
+
+- Body: 200-300 words built around the top 2-3 atoms
+- Include 1 pull quote (can reuse 1 from blog)
+
+- Clear CTA button text (e.g., "Listen Now", "Hear the Full Conversation")
+
+### Step 7: Pull Quotes
+
+Select 5-8 standalone quotes that:
+
+- Make sense without context
+- Are surprising, insightful, or emotionally resonant
+
+- Are under 150 characters each (for shareable graphics)
+- Include speaker attribution
+
+- Have timestamps (for linking to the exact moment)
+
+### Step 8: Quote Cards
+
+Generate visual quote card images for the top 3-5 pull quotes.
+
+**Requires Tier 2 dependencies** (Pillow + system fonts). Run `scripts/generate_quote_cards.py`:
+
+```python
+
+from generate_quote_cards import batch_generate
+
+quotes = [
+
+{"text": "Quote text here", "speaker_name": "Speaker", "speaker_title": "Title"},
+
+...
+
+]
+
+paths = batch_generate(quotes, output_dir="quote-cards/", podcast_name="Podcast Name")
+
+```
+
+Each quote produces 3 files:
+
+- `quote-N-instagram-1080x1080.png` (feed posts, carousels)
+- `quote-N-twitter-1200x675.png` (tweets, LinkedIn)
+
+- `quote-N-stories-1080x1920.png` (Instagram Stories, TikTok)
+
+Read `quote-cards.md` for design specs, color customization, and AI background generation.
+
+### Step 9: Short-Form Video Clips (Opus Clips Style)
+
+Automatically identify and extract the best 30-90 second moments from the episode for YouTube Shorts, TikTok, and Instagram Reels.
+
+**Requires Tier 3 dependencies** (ffmpeg + yt-dlp). For animated captions, also requires Tier 4 (OpenAI Whisper).
+
+The pipeline:
+
+1. Use content atoms (especially controversial takes, stories, and data points) to identify clip-worthy moments. **Atoms MUST have start/end timestamps** — atoms without timestamps are skipped.
+2. Download the source video if from YouTube (`scripts/generate_video_clips.download_youtube_video`).
+
+3. Trim clips at the identified timestamps with fade in/out.
+4. Reformat to 9:16 vertical (center-crop or letterbox).
+
+5. **If Tier 4 available:** Add animated word-by-word captions (Opus Clips style) via Whisper re-transcription of each clip.
+6. **If Tier 4 unavailable:** Add static segment-level captions using transcript data (still readable, just not word-highlighted).
+
+7. Burn in a progress bar and speaker name lower third.
+
+Target: 3-5 clips per episode, each scored by viral potential. Generate a hook title for each clip.
+
+```python
+
+from generate_video_clips import extract_clips
+
+results = extract_clips(
+
+video_path="episode.mp4",
+
+atoms=scored_atoms, \# must have start_timestamp/end_timestamp
+
+output_dir="clips/",
+
+speaker_name="Speaker Name",
+
+max_clips=5,
+
+)
+
+```
+
+Read `video-clips.md` for the full video clip extraction and captioning pipeline.
+
+### Step 10: Content Calendar
+
+Auto-schedule all generated content across platforms using best-practice timing:
+
+| Platform | Best Times | Frequency | Content Types |
+
+|----------|-----------|-----------|--------------|
+
+| Twitter/X | 9-11am, 1-3pm weekdays | 1-2 posts/day | Threads, quote cards, clip links, topic tweets |
+
+| LinkedIn | Tue-Thu 8-10am | 2-3 posts/week | Articles, quote cards, insights |
+
+| Instagram | Mon/Wed/Fri 11am-1pm | 3-5 posts/week | Quote cards, carousels, Reels |
+
+| TikTok | Tue/Thu/Sat 7-9pm | 3-5 posts/week | Short-form clips |
+
+| YouTube | Sat/Sun 9am-12pm | 1-2 Shorts/week | Short-form clips |
+
+| Newsletter | Tue/Thu morning | 1/week | Newsletter copy |
+
+| Blog | Any weekday | 1/episode | Blog post |
+
+Read `content-calendar.md` for scheduling logic, content mix optimization, and calendar output format.
+
+## Quality Checklist
+
+Run this checklist before delivering content to the user:
+
+- [ ] **No quote recycling**: No verbatim quote appears in more than 2 content pieces
+- [ ] **Platform leads are unique**: Twitter, LinkedIn, and blog each open with a different atom/angle
+
+- [ ] **Twitter character limits**: Every tweet is under 280 characters (count them)
+- [ ] **Blog word count**: Blog post is 800-1,200 words (count them)
+
+- [ ] **LinkedIn character count**: Post is 1,200-1,500 characters
+- [ ] **Instagram length**: Caption is 300-500 characters before hashtags
+
+- [ ] **Newsletter subject line**: Under 50 characters
+- [ ] **All links captured**: Show notes include every link from the YouTube description/episode metadata
+
+- [ ] **Timestamps on atoms**: Every atom has start*timestamp and end*timestamp
+- [ ] **Pull quotes are short**: Each under 150 characters
+
+- [ ] **Content pieces generated individually**: Each piece was generated in its own LLM call (not batched with ---SEPARATOR--- patterns that cause content to bleed between files)
+
+## Generation Best Practices
+
+**Generate each content piece in a separate LLM call.** Do NOT batch multiple outputs (e.g., "generate blog + social + newsletter separated by ---SEPARATOR---"). This causes:
+
+- Content bleeding between files (e.g., Twitter thread ends up in LinkedIn file)
+- Quality degradation on later pieces as the model runs out of attention
+
+- Inconsistent formatting
+
+Instead: generate atoms first, then generate each content piece independently with the atoms + relevant assignment as context. This produces better quality and avoids file-splitting bugs.
+
+**For long transcripts (>50K chars):** Split into 2-3 chunks with 5K char overlap. Extract atoms from each chunk. Merge and deduplicate before scoring.
+
+## Output Structure
+
+```text
+
+podcast-content/
+
+├── episode-title/
+
+│ ├── atoms.json \# Extracted + scored content atoms with timestamps
+
+│ ├── show-notes.md
+
+│ ├── blog-post.md
+
+│ ├── social/
+
+│ │ ├── twitter-thread.md
+
+│ │ ├── linkedin-post.md
+
+│ │ ├── instagram-caption.md
+
+│ │ └── topic-tweets.md \# (multi-topic episodes only)
+
+│ ├── newsletter.md
+
+│ ├── pull-quotes.md
+
+│ ├── quote-cards/ \# (Tier 2)
+
+│ │ ├── quote-1-instagram-1080x1080.png
+
+│ │ ├── quote-1-twitter-1200x675.png
+
+│ │ ├── quote-1-stories-1080x1920.png
+
+│ │ └── ...
+
+│ ├── clips/ \# (Tier 3+)
+
+│ │ ├── clip-1-vertical.mp4
+
+│ │ ├── clip-2-vertical.mp4
+
+│ │ └── clips-metadata.json
+
+│ ├── audiograms/ \# (audio-only or requested)
+
+│ │ ├── clip-1.mp4
+
+│ │ └── ...
+
+│ └── calendar.json \# Scheduled content calendar
+
+```
+
+## Tone & Voice Calibration
+
+Before generating content, ask the user (or infer from their existing content):
+
+- **Brand voice** — formal, conversational, edgy, academic, playful?
+- **Target audience** — who reads/follows their content?
+
+- **Platform priorities** — which platforms matter most?
+- **Minimum viral score** — threshold for content to make the cut (default: 40)
+
+If not specified, default to conversational and professional. Mirror the energy of the podcast itself — if the hosts are casual and funny, the marketing content should reflect that.
+
+## Composability
+
+This skill works alongside the `podcast-generator` skill. A typical workflow might be:
+
+1. Use `podcast-generator` to create an episode from research
+2. Use `podcast-marketing` to repurpose that episode into content
+
+## References
+
+- `content-atoms.md` — Content atom extraction prompt, the 7 atom types, output format, and scoring rubric. Read first — atoms are the foundation for all content.
+- `content-templates.md` — Detailed templates and examples for each content type (social posts, blog, newsletter, show notes, pull quotes). Read when generating specific content types.
+- `quote-cards.md` — Quote card design specifications, image generation, and multi-format output. Read when generating quote cards.
+- `video-clips.md` — Opus Clips-style video clip extraction, vertical reformatting, animated captions, and progress bar overlay. Read when generating short-form video clips.
+- `content-calendar.md` — Platform-specific scheduling rules, content mix optimization, and calendar JSON output format. Read when generating a content calendar.
+- `audio-processing.md` — Whisper transcription setup, audiogram generation pipeline, and audio clip extraction. Read when handling audio input or generating audiograms.
+- `rss-and-batch.md` — RSS feed parsing, episode downloading, and batch processing logic. Read when the input is an RSS feed or batch mode is requested.
+- `youtube-processing.md` — YouTube URL handling, caption extraction, audio download, and metadata retrieval. Read when the input is a YouTube link.