AI_Workflow

Knowledge Bases

The third layer of context: knowledge that doesn’t belong in any single repo, but that you want Claude to know before you start working. Frameworks, best practices, expert transcripts, style guides from people smarter than you.

This is what Claude Projects are for — and it’s a different tool from Claude Code. Worth understanding both.


Where each layer lives

Layer Tool What it’s for
1. Knowledge Claude Projects (claude.ai, the web app) Background knowledge that applies to many projects
2. Instructions CLAUDE.md (in your repo) Rules specific to THIS project
3. Operations Slash commands (.claude/commands/*.md) Repeatable verbs you’ll use repeatedly

Claude Code (the CLI you’ve been using) handles layers 2 and 3. Claude Projects handles layer 1.

What is a Claude Project?

On claude.ai, you can create a “Project” — a persistent conversation workspace with:

Use it for things like:

It is not for code-specific rules (those go in CLAUDE.md). It is for reference material that makes Claude smarter about a topic.

The TranScripts pattern

TurtleWolfe maintains ~/repos/TranScripts/, a library of cleaned YouTube transcripts organized by topic:

TranScripts/
├── Claude/
│   ├── Patterns/      # Claude Code workflow patterns
│   ├── Skills/        # Claude Code skills tutorials
│   └── Tools/         # Hooks, output styles, status lines
├── Docker/
│   └── Docker_Edited/ # Bret Fisher DockerCon talks, cleaned
├── Career/
│   ├── LinkedIn_Edited/   # LinkedIn profile frameworks
│   └── Resume_Edited/     # Resume writing frameworks
└── Drupal/
    └── Drupal_Edited/ # WebWash Drupal CMS tutorials

Each *_Edited/ folder:

  1. Contains YouTube transcripts stripped of filler (um, uh, intros, off-topic digressions)
  2. Keeps frameworks, actionable advice, templates, statistics
  3. Has a *_SYSTEM_PROMPT.md file that becomes the Claude Project’s instructions

The setup: on claude.ai, create a new Project. Paste the *_SYSTEM_PROMPT.md as the Project Instructions. Upload the entire *_Edited/ folder as the knowledge base. Now every conversation in that Project has expert knowledge baked in.

Why this is valuable

Without a knowledge base, every time you ask Claude about Dockerfiles you get generic advice. With a knowledge base full of Bret Fisher’s DockerCon talks, you get Bret Fisher’s opinionated specific advice. Claude cites the transcripts. The quality of answers jumps dramatically.

Same for resume writing. Same for LinkedIn. Same for any domain where there’s a specific expert or framework you want to emulate.

How to build a knowledge base

The recipe TurtleWolfe uses:

  1. Find a subject expert on YouTube you want to learn from
  2. Extract their transcripts (via mcp-youtube-transcript or manually)
  3. Clean them — strip filler, keep frameworks. TurtleWolfe has a /clean-transcript slash command that does this automatically
  4. Organize by topic in a repo
  5. Write a system prompt that tells Claude to cite the transcripts when answering
  6. Upload to a Claude Project at claude.ai

Full workflow in ~/repos/TranScripts/README.md if you want to steal it.

When NOT to use a knowledge base

Claude Projects vs Claude Code — which to use?

You’re… Use
Writing code in a specific repo Claude Code with CLAUDE.md
Asking questions about a topic (career, Docker, a framework) Claude Projects with a knowledge base
Doing both — writing code AND need expert background knowledge Both — Claude Code for the repo, a Claude Project open in a browser tab for reference

TurtleWolfe uses both, often in parallel. Code in Claude Code, tab over to Claude Projects to ask “what does Bret Fisher say about multi-stage builds?”, paste the answer back into Claude Code as context for the next prompt.