Claude Code can delete your whole repo in one command..π³
The .git, commits, branchs, all gone
The 30 sec fix in .claude/settings.json:
"permissions": {
"deny": ["Bash(rm -rf:*)", "Bash(git push --force:*)"]
}
Now those get blocked even if it tries. Still push your work often tho!
might be common knowledge but just in case:
you can run a coding AI on your own laptop FOR FREE without API keys or subscription AND your code never leaves your machine!
1. install Ollama
2. run: ollama run qwen2.5-coder
its def slower than the big models but great for quick stuff!:)
Always check these 4 things before you ship:
1. no API keys in the code
2. every route checks who is logged in
3. user input never goes straight into SQL
4. your .env is in .gitignore
stop writing a different instructions file for every coding tool!
AGENTS.md is one file that codex, cursor and github copilot all read
if you use claude code too, it reads CLAUDE.md so put this one line in it:
@AGENTS.md
Cursor can read your .env file..π³
their own staff confirmed the agent can read it even tho .env is skipped for indexing
fix: just add this to a .cursorignore file in your project:
.env*
still not bulletproof (terminal commands can read it)
just keep real secrets out of the project folder
sunday tip:
before you let ANY agent touch your code:
> git switch -c ai-try β work on a branch not main
> git diff β see what it changed
> git restore . β undo its edits to your files
new files it made stay put, just delete those yourself
if you vibecoded an app on Supabase, check this today:
> your anon key sits in your frontend code, anyone can find it
> if a table has row level security OFF, that key can read and write to it
FIX:
alter table your_table enable row level security
then add policies for what your app actually needs
you closed your terminal in the middle of a Claude code task?
claude -c β continue your last conversation
claude --resume β pick any older one from a list
the conversation is saved, just tell it to pick up where it left off:)
if you running claude with --dangerously-skip-permissions on your main pc, one bad command and it can touch anything you can..
use that mode inside a container or a throwaway vm, never on the machine with your keys and photos!
unpopular opinion: your claude md file is too long
it loads into every session. every rule competes with the actual task
i'd keep only the rules it actually breaks and delete the rest
before you let ai refactor anything, paste this first:
"Before changing any code, tell me what this code does today, the 3 riskiest parts and which tests will prove it still works after. Then stop and wait for me"
now its a plan you can check
if you vibecode an app with your OpenAi key or stripe secret key in it, anyone can open devtools and copy it if the key is in your frontend code!
keep keys on a server, the browser should only ever talk to your own api
every mcp server you add to your coding agent runs with YOUR permissions
it can read your files and run whatever its tools allow
before you install one make sure you know:
- who made it
- is the code public
- what tools does it add
and to see whats already installed:
claude mcp list