GUIDES
How to Build an Expert Knowledge File for Claude: Complete Prompt Guide đź“š
Your AI is pulling from the same sources as everyone else — mostly Reddit. Build a private knowledge file from premium Substacks, podcast transcripts, and X.
Where Your AI Gets Its Info
Most people don’t actually know where ChatGPT or Claude pulls their answers from. The full picture: they were trained on billions of pages of the public internet (Wikipedia, Reddit, news, blogs — basically anything free to access). When they need newer info, they do a live web search and read the top results.
The single biggest source they pull from is Reddit. And the implications are bigger than people realize.
Why Everyone Is Getting the Same AI Advice
Reddit signed a multi-year content-licensing deal with OpenAI in May 2024. Terms were never officially disclosed, but analysts estimate it’s worth roughly $60–70 million per year based on Reddit’s data-licensing revenue disclosures. That deal is what made Reddit the central training source it is today.
Independent research from 5WPR’s 2026 AI Platform Citation Source Index found Reddit accounts for roughly 40% of all citations across major AI search platforms — the highest of any single domain. ChatGPT briefly peaked near 60% Reddit citation share mid-2025 before settling lower.
So every time you ask ChatGPT or Claude for advice, the answer is heavily shaped by what’s upvoted on Reddit. Which means everyone using AI for advice is getting the same flavor of Reddit-derived answer.
Build Your Own Expert Knowledge File
The fix is straightforward: build a private knowledge file Claude can read every time you work together. Fill it with sources that AREN’T scraped by the major training runs — podcast transcripts, premium Substacks, expert X threads, niche industry newsletters — specific to YOUR field.
Now your AI is pulling from the people actually winning in your industry, not generic Reddit consensus. Here’s the exact setup I run.
For my e-commerce + AI work, mine are: 5-10 premium Substacks from operators I trust, 3-4 podcasts I listen to weekly, the 10-15 X accounts that post real signal in my space, and a couple of industry newsletters. Whatever your field, pick the sources where YOUR experts actually publish.
I use Apify as the data layer. It has off-the-shelf actors that scrape Substack posts, transcribe podcast audio (YouTube + Spotify), pull X threads, and grab newsletter archives. Set up one actor per source. Runs on a schedule.
In Claude Code, create a Project (or a Claude Code folder) called something like knowledge/. Each source gets a subdirectory: knowledge/substacks/, knowledge/podcasts/, knowledge/x-threads/. Apify outputs land in the right folder automatically.
In your Project Instructions, point Claude at the directory: “Before answering any question about [my industry], search the knowledge/ directory for relevant context first.” Claude will now pull from your private corpus before falling back to its training data. The answers shift immediately.
Set the Apify actors to run weekly so the knowledge stays current. The compounding part is real: after 3 months, you’ve got a private corpus of the smartest content in your field, all searchable by Claude. After 12 months, your AI is genuinely an expert in your space — not because you trained a model, but because you fed it the right reading list.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.