prompt-git (pgit)

Git for your prompts. Commit, diff, branch, merge, and rollback LLM prompts — with semantic diffs that tell you what meaning changed.

The problem

Prompts are load-bearing infrastructure in production AI systems — but they're managed like config files: no history, no diffs, no rollback.

When your agent regresses at 2 AM, you have no idea:

• Which prompt changed?
• What exactly changed in it?
• Was it broken before, or did a recent edit introduce the regression?

The solution

$ pgit diff HEAD~1 --semantic

  Semantic diff: 7b1e4f1 → a3f9c2d
  ───────────────────────────────────────────────────────────
  Summary:   Tone shifted from permissive to strict; added hard
             limits on off-topic queries.

  Added:     - Explicit instruction to refuse off-topic questions
             - Hard word limit of 200 tokens per response
             - Escalation path for complex queries

  Removed:   - "Feel free to explore" phrasing
             - Open-ended example in the few-shot section

  Tone:      Permissive → Strict / Professional

pgit gives prompts the same rigorous version control that code has had for 30 years — plus one thing code diffs can't do: describe what the meaning changed.

Install

# Could not get prompt-git project name :-( 
pip install sr-prompt-git

# For semantic diff (requires Anthropic API key):
pip install sr-prompt-git[semantic]
export PGIT_LLM_KEY=sk-ant-...

Quick start

# 1. Initialise a prompt repo in your project
pgit init

# 2. Stage and commit your first prompt
echo "You are a helpful assistant. Be concise." > system.md
pgit add system.md
pgit commit -m "initial system prompt"

# 3. Attach your eval scores to the commit
pgit eval attach pass_rate 0.87
pgit eval attach avg_cost_usd 0.003

# 4. Experiment on a branch
pgit branch create experiment/stricter
pgit branch switch experiment/stricter

echo "You are a strict assistant. Refuse all off-topic queries." > system.md
pgit add system.md
pgit commit -m "tightened safety boundaries"
pgit eval attach pass_rate 0.94

# 5. See what semantically changed
pgit diff HEAD~1 --semantic

# 6. Merge back — only if pass_rate improved by 5%+
pgit branch switch main
pgit merge experiment/stricter --if-better pass_rate
#  ✓ Merged: pass_rate 0.87 → 0.94 (+8.0%)

Why prompt-git?

Full CLI reference

pgit init                                      Initialise a prompt repo
pgit add <file> [--model <hint>]               Stage files for commit
pgit commit -m <msg> [--author <name>]         Commit staged files
pgit status                                    Show staged / unstaged state
pgit log [--branch <b>] [--limit N] [--json]  Commit history with eval scores
pgit diff [<from>] [<to>] [--semantic]         Line diff or LLM semantic diff
pgit branch list                               List branches
pgit branch create <name> [--from <ref>]       Create a branch
pgit branch switch <name>                      Switch branches
pgit branch delete <name>                      Delete a branch
pgit merge <branch> [--if-better <metric>]     Merge (optionally gated on eval)
pgit tag <name> [<commit>] [-m <msg>]          Tag a commit
pgit eval attach <metric> <value>              Attach eval score to HEAD
pgit eval show [--commit <hash>] [--json]      Show eval scores
pgit push [<remote>] [<branch>]                Push to remote
pgit pull [<remote>] [<branch>]                Pull from remote

<from> / <to> accept: commit hashes, branch names, tag names, HEAD, HEAD~N.

Supported prompt formats

How it works

Everything lives in .promptgit/store.db — a plain SQLite file you can open with any viewer.

Objects are content-addressed (SHA-256, like git):

• Blob — immutable prompt text; identical content stored once regardless of commits
• Tree — maps file paths → blob hashes
• Commit — tree + parent + message + author + timestamp; immutable once written
• Tag — named pointer to a commit
• SemanticDiff — LLM-generated meaning diff, cached permanently by (from_hash, to_hash)

Eval scores live in a separate table linked by commit hash — so attaching scores never changes a commit's hash.

Semantic diff

pgit diff HEAD~1 --semantic            # all changed prompts, HEAD~1..HEAD
pgit diff v1.0 v2.0 --semantic --file system.md   # one file, specific tags
pgit diff --semantic --json            # machine-readable JSON output

• Powered by claude-haiku-4-5-20251001 by default (fast, cheap)
• Override model: PGIT_LLM_MODEL=claude-sonnet-4-6
• Results are cached permanently — same pair of versions never hits the LLM twice
• Always opt-in via --semantic — never called automatically

Environment variables

TypeScript SDK

A read-only TypeScript SDK is included for teams reading prompt history from Node.js:

import { PromptRepo } from 'prompt-git'

const repo = PromptRepo.open('/path/to/project')
const commits = repo.log({ limit: 10 })
console.log(commits[0].message)

Examples

See the examples/ directory:

• basic_workflow.sh — init → add → commit → diff
• team_sync.sh — push/pull with a shared remote
• eval_driven_merge.sh — eval-gated merges in CI

Development

git clone https://github.com/your-org/prompt-git
cd prompt-git
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

pytest tests/ -v        # run tests
mypy pgit/ --strict     # type check
ruff check pgit/        # lint

Tests use cassette recordings — no live LLM calls in CI.

Contributing

See CONTRIBUTING.md. New parsers, remote backends, bug fixes, and docs all welcome.

License

MIT © prompt-git contributors