I Built Git for AI Prompts — Here's Why and How
Every engineer I know who builds with LLMs has the same problem.
You spend hours tuning a system prompt. It's working great. You tweak it a little. Then a little more. Then something breaks — the model starts giving worse answers, hallucinating more, ignoring your instructions.
And you have absolutely no idea what you changed.
Your prompt history is scattered across files, Notion docs, Slack messages, and git commits buried inside application code. There's no clean way to see what changed, when, or why — let alone get back to the version that was actually working.
So I built promptctl.
What is it?
promptctl is a CLI tool that brings the git mental model to prompt management. You commit versions, diff them, roll back, search across them — all from the terminal.
$ echo "You are a helpful assistant." | promptctl commit system -m "initial"
✓ Committed prompt "system" as v1
$ echo "You are a helpful assistant. Always cite sources." | promptctl commit system -m "added citation"
✓ Committed prompt "system" as v2
$ promptctl diff system
--- system v1 (2026-06-01 09:12)
+++ system v2 (2026-06-03 14:47)
- You are a helpful assistant.
+ You are a helpful assistant. Always cite sources.
$ promptctl rollback system 1 -m "citation hurt recall"
✓ Rolled back "system" to v1 → saved as v3
If you've used git, you already know how to use promptctl.
The full workflow
Committing a prompt
You can pipe from stdin, read from a file, or type interactively:
# From stdin
echo "You are a helpful assistant." | promptctl commit system -m "initial"
# From a file — best for longer prompts
promptctl commit system --file prompts/system.txt -m "from file"
# Tag with the model it was written for
promptctl commit classifier -m "optimized for speed" --model gpt-4o-mini --tag prod
Viewing history
$ promptctl log system --preview
prompt: system
──────────────────────────────────────────────────
v3 2026-06-24 19:41:34
citation hurt recall
1 lines, 5 words, 28 chars
"You are a helpful assistant."
v2 2026-06-24 19:41:25
added citation
1 lines, 8 words, 49 chars
"You are a helpful assistant. Always cite sources."
v1 2026-06-24 19:41:25
initial
1 lines, 5 words, 28 chars
"You are a helpful assistant."
Diffing versions
promptctl diff system # latest vs previous
promptctl diff system 2 # v2 vs latest
promptctl diff system 1 3 # explicit comparison
Diffs are colorized — red for removed lines, green for added.
Searching across everything
$ promptctl search "cite sources"
Results for "cite sources"
──────────────────────────────────────────────────
system v2 2026-06-24 19:41
added citation
…You are a helpful assistant. Always cite source…
Search checks content, commit messages, and tags.
Watching a file
This one's my favorite. Point it at a file and it auto-commits every time you save:
promptctl watch prompts/system.txt --as system --model claude-3
Now you can edit in your normal editor and every save is a versioned snapshot. The commit timestamp matches the file's modification time, not when promptctl ran.
Other commands
promptctl copy system system-experimental # fork with full history
promptctl show system --copy # copy to clipboard
promptctl show system --version-at 2026-06-01 # time travel
promptctl export system > history.md # full markdown export
promptctl stats # store-wide overview
promptctl prune system --keep 10 # housekeeping
How it works
Everything is stored in .promptctl/store.json in your project directory — similar to how .git/ works. The store is discovered by walking up parent directories, so commands work from any subdirectory.
your-project/
├── .promptctl/
│ └── store.json ← all prompt versions live here
├── src/
└── ...
Commit store.json to git and your whole team shares prompt history. Or add .promptctl/ to .gitignore if you want a local-only store.
Writes are atomic — we write to a temp file and rename, so you never get a corrupt store even if the process is killed mid-write.
Why Go, why zero dependencies?
I wanted promptctl to be a tool you install once and forget about. No runtime required, no node_modules, no pip install, no version conflicts.
Go's standard library covers everything promptctl needs:
- JSON serialization
- Atomic file writes
- Directory traversal
- Diff algorithm (LCS, implemented from scratch)
- ANSI color detection
The result is a single binary you install with:
go install github.com/naya-ai/promptctl/cmd/promptctl@latest
And it works on Mac, Linux, and Windows.
Shell completions
# Bash
source <(promptctl completion bash)
# Zsh
source <(promptctl completion zsh)
# Fish
promptctl completion fish > ~/.config/fish/completions/promptctl.fish
Completions dynamically suggest your prompt names for every command that takes a <name> argument.
What's next
This is v0.1.0 — the core workflow is solid but there's a lot more to build. Things I'm thinking about:
- Remote sync (S3, GitHub Gist) so prompts are backed up and shareable
- Side-by-side diff view
- VS Code extension
If you build with LLMs and prompt management is painful for you, I'd love to hear what's missing.
Try it
go install github.com/naya-ai/promptctl/cmd/promptctl@latest
promptctl init
echo "You are a helpful assistant." | promptctl commit system -m "initial"
promptctl log system
GitHub: github.com/naya-ai/promptctl
If this solves a problem you have, a star on the repo goes a long way for a solo project. And if something's broken or missing, open an issue — I respond fast.