prpack — Summary

prpack is a zero-dependency Node.js CLI that packs a pull request into a single markdown file optimized for LLM code review — containing the commit list, the per-file diff, and the full post-change content of every touched file. The author's core insight (documented in THE_TECHNIQUE.md) is that LLM code review fails not because of model quality but because the diff alone lacks the surrounding context the model needs to spot bugs in unchanged lines. prpack solves the context problem by appending the full post-change file contents alongside each diff section, ensuring the model sees "what the file looks like after the change" rather than only what changed. A --review flag directly calls the Anthropic Messages API and streams a focused review (security, performance, tests, architecture, or general) to the terminal. An optional GitHub Action wrapper (prpack-action) enables automatic packed context on every PR. Compared to seeds, prpack has the narrowest scope of all ten batch frameworks — it is a pure context-packaging utility with no behavioral framework, no agent orchestration, and no persistent state, most similar to codesight in intent (reduce token waste for a specific AI task) but different in domain (PR review vs. codebase exploration).

prpack — Overview

Origin

Created by Lucas2944. The technique behind prpack is documented in THE_TECHNIQUE.md — a story of a null-deref bug that Claude Code missed because it only saw the diff, not the surrounding function context. The README states: "This is not a Claude problem. This is a context problem."

Philosophy

From THE_TECHNIQUE.md:

"When you paste a diff into a model, you are asking it to reason about a change in isolation. Diffs are designed for humans who already have the file open in an editor. The human reviewer mentally fills in the surrounding code, the call sites, the types, the invariants. The model does none of this."

"packs exactly the diff plus the full state of every file the diff touches. That's enough context to reason about the change without drowning the model in noise."

"The hit rate on substantive feedback goes from roughly one in five to roughly four in five."

Scope

prpack is intentionally narrow: "It's a single CLI binary." The README explicitly documents what it does NOT do:

• No network calls in pack mode
• No hidden AI
• No web UI
• No telemetry

Ecosystem

• prpack-prompts — four focused review prompts (security / performance / tests / architecture) as plain markdown files, usable independently
• prpack-action — GitHub Action wrapper (5-line workflow YAML)
• prpack Pro Pack — curated .prpack.yml configs for specific review styles (optional, free/PWYW on itch.io)
• prpack-demo — browser-side demo at lucas2944.github.io/prpack-demo

Key Design Decisions

1. Zero dependencies — requires only Node 18+ and git on PATH
2. Full file content beats targeted extraction — include the whole post-change file, not just adjacent lines
3. Configurable via YAML — .prpack.yml adds preface notes and custom review prompts without changing the CLI
4. Review angles are first-class — security, performance, tests, architecture are supported as discrete modes, not freeform prompts

prpack — Architecture

Distribution

• Type: npm package (global or npx)
• Binary: prpack (mapped from bin/prpack.js)
• Version analyzed: 0.2.0
• Runtime required: Node.js 18+, git on PATH
• Dependencies: Zero runtime dependencies
• Language: JavaScript

Install Methods

# One-shot, no install
npx github:Lucas2944/prpack --out ctx.md

# Global install
npm install -g github:Lucas2944/prpack
prpack --out ctx.md

Directory Tree

prpack/
├── bin/
│   └── prpack.js          # CLI entry point (shebang)
├── src/
│   ├── config.js          # Config loading (.prpack.yml)
│   ├── git.js             # Git operations (diff, log, file content)
│   ├── pack.js            # Core packing logic
│   ├── prompts.js         # Review prompt templates
│   └── review.js          # Anthropic API integration
├── schema/
│   └── prpack.schema.json # JSON Schema for .prpack.yml
├── examples/              # Side-by-side comparison examples
├── test/
├── THE_TECHNIQUE.md       # Extended philosophy document
├── CHANGELOG.md
└── package.json

Config File

.prpack.yml (project-level, optional):

base: develop
includeTests: true
exclude:
  - dist/**
  - "*.lock"
preface: |
  This PR introduces a new payment flow. Focus on input validation.
reviewPrompt: |
  You are a senior engineer reviewing this PR. Flag bugs...

GitHub Action

# .github/workflows/prpack.yml (5-line integration)
- uses: Lucas2944/prpack-action@v1
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
    review_angle: general

Target AI Tools

• Claude (primary — has native --review integration)
• Cursor
• Any LLM accepting markdown context (paste ctx.md)

prpack — Components

CLI Options

Review Angles

Source Modules

Output Shape

# Pull Request Context
**Repo:** `org/repo`
**Branch:** `feature/...`
**Base:** `main` → **Head:** `HEAD`
**Commits:** 3
**Files changed:** 7

## Commits
- `abc1234` 2026-05-10 — Add Stripe webhook handler

## Files changed
- `src/payments/webhook.ts` _(added)_

---
## `src/payments/webhook.ts` _(added)_

### Diff
```diff
...

Full content (post-change)

...

## No Hooks, No Skills, No MCP

prpack ships zero hooks, skills, subagents, or MCP servers. It is a pure CLI tool.

prpack — Uniqueness

differs_from_seeds

prpack has no direct analog among the 11 seeds — none of them address PR-level context packaging for LLM review. The closest parallel is codesight (in this batch), which solves the analogous "session-start context" problem by pre-compiling codebase structure. prpack solves the "PR review context" problem by packaging git diff + full file contents. Both are pre-session static packagers with zero behavioral framework components. Among seeds, agent-os is the most philosophically similar (produce files that agents read) but agent-os focuses on long-lived project conventions while prpack is ephemeral (one file per PR). Unlike spec-kit, claude-conductor, or superpowers, prpack has no skills, hooks, commands, or agent workflow instructions whatsoever — it is purely a context formatter.

Positioning

• Primary differentiator: The only framework in the batch that targets PR review specifically, and the only one with a formal manifesto (THE_TECHNIQUE.md) explaining the context engineering insight behind the tool.
• Secondary differentiator: Zero dependencies is unusual for a Node.js CLI — no bundled LLM SDK at install time; the Anthropic SDK is called only if --review is used, and only the ANTHROPIC_API_KEY env var is needed.
• Target user: Developers who do code review with Claude/LLMs and are frustrated by missed bugs in unchanged context.

Observable Failure Modes

1. Large PRs hit token limits: Very large PRs (many files, large files) will produce ctx.md that exceeds model context windows. --no-content and --max-bytes mitigate this but degrade review quality.
2. Generated files bloat output: Without careful --exclude configuration, build artifacts and lock files inflate ctx.md significantly.
3. Review mode is Anthropic-only: --review requires ANTHROPIC_API_KEY — no provider abstraction, no local model support.
4. No semantic understanding: prpack is entirely syntactic — it does not understand code semantics, so file selection is purely "files touched by the diff" rather than "files relevant to understanding the change."
5. Stars indicate early-stage adoption: 2 stars (as of analysis date) suggests very early community adoption despite a well-articulated problem statement.