Three-Tool Pipeline Example: Scrape, Transcribe, Publish

The skill soup thesis is that combining narrow tools beats one giant tool. Last week I built a pipeline that proves the point: scrape a video page, transcribe the audio, publish a post. Three tools, one workflow, zero heroics. Here is how the pieces fit.

The Problem

I had a backlog of recorded talks that needed to become blog posts. Manually this is hours of work each: watch, transcribe, edit, format, publish. The skill soup approach is to let each tool handle the part it is good at.

The Pipeline

Step 1: fetch video page HTML         -> Python + httpx
Step 2: extract audio track           -> yt-dlp + ffmpeg
Step 3: transcribe audio              -> OpenAI Whisper
Step 4: clean and structure transcript -> Claude via API
Step 5: publish as draft post         -> local posts API

No tool does more than one job. Each one has a clear input and output. If any stage breaks I can rerun from the last good artifact without redoing the rest.

The Glue Code

The pipeline itself is about 80 lines of Python. The heavy lifting is in the tools; the glue just moves artifacts between them:

def pipeline(video_url: str) -> str:
    html = fetch_page(video_url)
    audio = extract_audio(video_url)
    transcript = whisper_transcribe(audio)
    structured = claude_structure(transcript)
    slug = publish_draft(structured)
    return slug

Each function is small. Each function has its own error handling. If Whisper is slow that day, only transcription is slow. If Claude hits a rate limit, I retry just that step.

What Makes This Work

The whole thing only works because every tool produces a persistable artifact. The HTML is a file. The audio is a file. The transcript is a file. The structured post is a file. That means the pipeline is restartable and cacheable. Running it again on the same input skips work that is already done.

This is the skill soup in practice. Not a single monolithic system that does everything. A set of narrow tools glued together by a workflow that respects each one's boundaries.

Read the OpenAI Whisper documentation for transcription options.