I Built Git for LLM Prompts — Because `summarize_v3_FINAL_actually_final.txt` Isn't Working

The Problem Nobody Talks About

I track everything in git. My code. My configs. Even my dotfiles. But my AI prompts? They lived in a messy folder called /prompts with files like:

summarize.txt
summarize_v2.txt
summarize_better.txt
summarize_FINAL.txt
summarize_FINAL_v2.txt
summarize_v3_FINAL_actually_final.txt

Sound familiar?

The problem has a name: prompt drift. You tune a prompt obsessively until it works. You ship it. It works great. Then 3 months later, you want to roll back — and you have no idea what changed, when, or why. You don't know which version you actually used in production. You don't know what temperature you had it at when it stopped hallucinating.

Regular git helps a little, but it doesn't understand prompts. It can't track which model you were using, what temperature you tested at, or attach performance notes to a specific version.

So I built prompt-git.

---

The Solution: 3-Line Demo

$ prompt-git init
$ prompt-git add summarize.txt
$ prompt-git commit -m "tighten output format" --model claude-3-5-sonnet --temp 0.3 --notes "BLEU score +12%"

That's it. Now you have a real version history for your prompts.

---

What prompt-git Actually Does

# See your full prompt history in a beautiful table
$ prompt-git log

┌──────────┬────────────┬──────────────────────────┬─────────────────────┬──────┬────────────┬──────────────────┐
│ ID       │ Prompt     │ Message                  │ Model               │ Temp │ Tags       │ Timestamp        │
├──────────┼────────────┼──────────────────────────┼─────────────────────┼──────┼────────────┼──────────────────┤
│ b7d9e441 │ summarize  │ tighten output format    │ claude-3-5-sonnet   │ 0.3  │ v1.0       │ 2026-02-24 10:15 │
│ a3f8c12e │ summarize  │ initial summarize prompt │ claude-3-5-sonnet   │ 0.7  │            │ 2026-02-24 09:00 │
└──────────┴────────────┴──────────────────────────┴─────────────────────┴──────┴────────────┴──────────────────┘

# Diff two versions to see exactly what changed
$ prompt-git diff a3f8c12e b7d9e441

--- summarize@a3f8c12e
+++ summarize@b7d9e441
-Summarize the following text in 3 bullet points:
+Summarize the following text in 3 concise bullet points.
+Focus on key facts only. Do not include opinions or speculation.

+2 additions  -1 deletions

# Tag your best version
$ prompt-git tag v1.0 -m "Production-ready, BLEU +12%"

# Restore any version instantly
$ prompt-git checkout v1.0 --prompt summarize --write

---

Architecture

The design is intentionally simple:

src/prompt_git/
├── cli.py      # Click commands + Rich terminal output
├── store.py    # SQLite backend (no external services)
├── diff.py     # Unified + word-level diff engine
└── models.py   # PromptCommit, PromptTag, PromptFile dataclasses

Everything lives in a .prompt-git/store.db SQLite file in your project directory. No API keys. No cloud. No dependencies beyond click and rich. It's just a fast local database that understands the shape of prompt version history.

Each PromptCommit stores:

• The full prompt text
• A SHA-1 content hash as its ID
• Parent commit ID (linked list of history)
• Optional: model, temperature, notes

This means you can track not just what changed, but why and with what settings — the context that git alone can't give you.

---

What Makes This Different from Just Using Git?

You might be thinking: "Why not just commit your prompts to git?"

You can! And you should. But prompt-git adds things that git doesn't know about:

Feature	Regular git	prompt-git
Content diff	Yes	Yes
Commit messages	Yes	Yes
Tags / releases	Yes	Yes
LLM model metadata	No	Yes
Temperature tracking	No	Yes
Performance notes	No	Yes
Prompt-scoped history	No	Yes
Works without a git repo	No	Yes
Rich terminal UI	No	Yes

The metadata is what makes it genuinely useful. When you see that v1.0 used temperature=0.3 and claude-3-5-sonnet and had a note saying "reduced hallucinations by 40%", you have actionable context — not just a diff.

---

What I Learned Building This

1. SQLite is wildly underrated for CLI tools. A single .db file gives you transactions, queries, foreign keys, and WAL mode for free. No server, no config. It's the perfect backend for a local developer tool.

2. Rich makes terminal UIs genuinely delightful. The colored diff output, bordered panels, and formatted tables took maybe 30 minutes to add but make the tool feel professional. Never use print() in a CLI tool again.

3. Click's CliRunner is incredibly good for testing. I wrote the entire test suite using CliRunner.isolated_filesystem() — each test gets its own temp directory, no mocking required. The tests are fast, readable, and they catch real bugs.

4. The hardest part wasn't the code — it was the UX. What does checkout mean for a prompt? Should it write to a file? Should it just print to stdout? Getting those interaction patterns right took more thought than the SQLite schema.

---

Install and Try It

pip install prompt-git
cd your-ai-project/
prompt-git init
prompt-git add your-prompt.txt
prompt-git commit -m "first tracked version"
prompt-git log

Full source: github.com/LakshmiSravyaVedantham/prompt-git

If you're building LLM applications and you're not tracking your prompts this way, your future self will thank you.

---

Built with Python, Click, Rich, and SQLite. No AI APIs required — this is version control, not inference.