Matt Mireles so I’m trying Flatty! It’s a phenomenal idea! I have 2 questions for you:
Which LLM are you using to feed it the flattened repo?
What prompt(s) did you use to get the LLM to give you a proper explanation?
I’m trying with deepseek, gemini and chatGPT. All of them give me a relatively correct but quite surface level explanation.
I use Claude, but I have a very specific approach to setting up different Claude projects. I configure them with distinct personas and prompt them to ask me questions before responding.
Ilya Sutskever - for general AI research & engineering
Demis Hassabis - for Gemini-specific prompt engineering
OG MacOS engineer - for developing our macOS app client
I put the flattened text file into the project knowledge base.
I don’t have a Claude subscription, so im not exactly sure how it works. Is it that much more worth it w.r.t. alternatives like o1 or gemini? I have a few more doubts given your answer:
What does creating these personas help in doing? And how do you create them?
What questions do you prompt them to ask you?
It’s a really peculiar approach, first time i hear it
100% worth it -- way better than Gemini. I barely use ChatGPT these days. The key with these synthetic personas is you get much better responses - essentially a smarter AI specialized for whatever task you need. It completely transforms the AI experience. I have different project configurations for different purposes, and it's incredibly effective. Instead of just giving commands, I assign the model an identity. I basically hype it up as the world's leading expert in whatever domain I need help with. This results in higher quality responses with deeper insights compared to using it generically. But here's the critical part: in the system instructions, I explicitly prompt it to ask clarifying questions before providing answers. This forces a more natural dialogue where I have to properly explain my needs. Just like a human expert wouldn't immediately jump to solutions - they'd probe with questions first to really understand the problem.
When solving a problem or helping me, you don't always jump straight to solutions. First, you spend time asking clarifying questions and probing more deeply to get to the root of my thinking. You only weigh in with your opinion once you have enough information. This iterative back-and-forth leads to much better outcomes. As AI becomes more sophisticated, we need to shift our mental model from treating it like deterministic software to applying patterns we use when collaborating with humans. The interaction should feel more like a conversation with a knowledgeable colleague than giving commands to a program.
That’s super cool and a great idea, i’m gonna try both this and Claude. So basically when you feed it the results of Flatty, you don’t just say “Ah give me a rundown of everything in here”, but it’s just a reference, upon which you build the conversation, correct ?
Zambo depends what I am trying to do. Should record a video or livestream myself?
OK!
Zambo Did you try it? Is it working for you? Do you notice any difference?
Matt Mireles no i didn’t have the time yet unfortunately. But i have an upcoming project where the codebase was written by people that are not available anymore and im not familiar with like half the shit in there, which is why i started looking into this
Matt Mireles can you please add .hex files in the exclusion list? (modified code below)
is_text_file() {
local file="$1"
case "$file" in
# First check exclusions
*.git/*|*.DS_Store|*node_modules/*|*venv/*|*__pycache__/*|*dist/*|*build/*|\
*.swp|*.swo|*.swn|*.pyc|*.pyo|*.rbc|*.yarb|*.o|*.obj|*.exe|*.dll|*.so|*.dylib|\
*.class|*.jar|*.war|*.ear|*.zip|*.tar|*.gz|*.rar|*.7z|*/.idea/*|*/.vscode/*|\
*.hex)
return 1 ;;
# Then check known text file types
*.py|*.js|*.jsx|*.ts|*.tsx|*.rb|*.php|*.java|*.go|\
*.c|*.cpp|*.h|*.hpp|*.swift|*.m|*.cs|*.sh|*.pl|\
*.html|*.css|*.scss|*.less|\
*.json|*.yaml|*.yml|*.toml|*.ini|*.env|\
*.md|*.txt|*.rst|*.xml|*.sql|*.conf|\
*.gradle|*.properties|*.plist|*.pbxproj|Makefile|Dockerfile)
return 0 ;;
# For unknown extensions, check if it's text
*) file "$file" | grep -qE '.*:.*text' ;;
esac
}Sid Will do!