AI Content Pipeline

First things first: what's a .md file?

A .md file is a Markdown file. It's just a plain text file with some simple formatting (headings, bold, lists) that any text editor can open. Think of it as a step between a plain .txt file and a Word document.

They matter here because Claude Code reads .md files natively. Every skill, every reference doc, every log file in this system is a .md file. They're lightweight, easy to edit, and Claude understands them perfectly.

Your reference docs

Download these two files. They're the detailed knowledge base behind this guide.

Where to put them

Drop both files into a docs/ folder in your project:

your-project/
  docs/
    fal-video-pipeline-guide.md
    workflow-notes.md

Why this matters

Skills work best when they're short and focused. Instead of cramming everything into the skill file, you point the skill at these docs for the detail. Claude reads them when it needs context and ignores them when it doesn't.

How to reference them from a skill

Add this line to any skill that needs the technical detail:

Read docs/fal-video-pipeline-guide.md for API setup, prompt file format, and runner script details.

Or for the workflow context:

Read docs/workflow-notes.md for the full pipeline overview and production requirements.

Tools in the mix

The pipeline

Each step below becomes its own skill. You run them in order, checking the output at each stage.

What's a skill?

A skill is a .md file called SKILL.md saved in your project. It tells Claude exactly what to do for a specific task.

You trigger it by typing its name as a slash command:

/generate-images

Skills can reference scripts, docs, and each other. They only run when you tell them to. No surprise API calls.

Skills you'd build

Making skills better over time

This is the part that stops you repeating the same mistakes.

The loop

Add this to every skill

Put this block at the end of every SKILL.md you create. It tells Claude to log failures so you can fix the skill later:

## Logging

When any step fails or produces a result that doesn't match the brief:
1. Log what happened to docs/skill-logs.md
2. Include: which step failed, what the input was, what went wrong, and a one-line suggestion for fixing it
3. Keep log entries short. Just enough context to debug, not a wall of text.
4. Do not stop the rest of the batch for a single failure.

Review logs in a new session

Don't try to fix a skill in the same session that broke it. Open a fresh Claude Code session, point it at the log, and say:

Read docs/skill-logs.md and improve the [skill name] skill based on what failed.

Keep skills focused

Each skill should do one thing well. If it needs detailed context, point it at a file in docs/ rather than cramming it all into the skill itself. Less context means faster, more accurate results.

Stay current

Prompting techniques and model capabilities change fast. Use the 30-day search skill regularly to check for better approaches before locking in a new workflow.

Keeping costs down

Quality checks

Add this block to any image or video generation skill:

## Quality checks

Before marking a generated image as approved:
1. Check the image matches the original brief.
2. Verify the player's face looks like the right person.
3. Check jersey numbers match the current squad list.
4. Flag any images that appear to use copyrighted source photos. Set these aside for human review.
5. If a check fails, retry up to 2 times before logging the failure and moving on.

Getting started

Set up Claude Desktop

• Open Claude Desktop and go to the Code tab.
• Point it at your project folder.
• Start in "Ask permissions" mode so you approve each action while learning.

Add your API keys

Add these in the local environment editor inside the Code tab (not in any file):

FAL_KEY=your-fal-key-here
ELEVENLABS_API_KEY=your-elevenlabs-key-here

Build one skill at a time

Get it working with 1 or 2 test jobs. Then move to the next. Once you trust a skill, relax the permissions so it runs with less hand-holding.

Rules to live by