🤖 Generic AI hacking workflow

I’ve been using a few different models while doing coding work, and over xmas I finalized a workflow that works with them all, excluding copilot which only works in an editor. So far this technique works well with:

  1. The consultancy’s internal model
  2. The client’s multiple internal models, and external ones
  3. ChatGPT, using my account.
  4. ollama, running locally or on a rented machine on vast.ai over ssh (-X)
  5. claude
  6. grock
  7. groq
  8. gemini

So I’ve added aliases for c, v and git-dump, to my ~/.bashrc and/or ~/.zshrc files:

if command -v pbcopy &>/dev/null; then
    alias c="pbcopy"
    alias v="pbpaste"
elif [ "$XDG_SESSION_TYPE" = "wayland" ] && command -v wl-copy &>/dev/null; then
    alias c="wl-copy"
    alias v="wl-paste"
elif command -v xsel &>/dev/null; then
    alias c="xsel --clipboard --input"
    alias v="xsel --clipboard --output"
elif command -v clip.exe &>/dev/null; then
    alias c="clip.exe"
    alias v="powershell.exe Get-Clipboard"
fi


# commit hash of an empty git repo
EMPTY_TREE=$(echo -n 'tree 0\0' | sha1sum | cut -f 1 -d ' ')

alias git-dump="git diff $EMPTY_TREE --"

Then when I want to start a conversation about a project or subset of it, I can run git-dump | c and paste it in from the clipboard. Like:

Hey chatbot, here's where we're up to. Summary and next steps please:

<ctrl+v>

Then it gets all the files as a single patch, but doesn’t have the full commit history which will blow the context window or give too much comprognitive load; it works with reasonably strong models.

The most powerful thing about this, is that adding instructions for an AI agent to the project’s README.md will cause it to focus on the next tasks rather than noise that’s less relevant.

If it’s a large project or I want to focus on just a small area I can filter by a dir, like git-dump . | c, add filters to exclude binary files or large text files like so: git-dump . ':!*.cast'.

And when the model gives some output, I can v > file to write the changes, and if I make my own changes and feed them back in, just cat file | c to share my own edits.

And of course it works with other things like git-dump | wc -l, with grep, or v | pastebinit and so on.

Given how fast this space is moving, it might not remain the best way to do things for very long. But for the moment, it works pretty well for me.