sourcebook

Your codebase has conventions your AI agents keep missing.

INITIALIZE_ENVIRONMENT arrow_forward

$ npx sourcebook init --format all

check_circle Scanned 10,456 files, 3 frameworks detected

check_circle Extracted 11 findings

● Core modules: types.ts imported by 183 files

● Circular deps: bookingScenario.ts ↔ getMockRequestData.ts

● Co-change: auth/provider.ts ↔ middleware/session.ts (88%)

● Dead code: 1,907 orphan files detected

DONE Wrote CLAUDE.md, .cursorrules, copilot-instructions.md

3.1s — only non-discoverable information.

ALGORITHM_01

Import Graph PageRank

Ranks every file by structural importance. Find the hubs that break everything when you touch them.

Hub: types.ts (183 importers)
Hub: fixtures.ts (86 importers)
Hub: http.ts (72 importers)

ALGORITHM_02

Git Forensics

Reverted commits are "don't do this" signals. Co-change coupling reveals invisible dependencies.

CO-CHANGE: auth.ts ↔ session.ts
CORRELATION: 88%
REVERTS: 2 found (anti-patterns)

ALGORITHM_03

Convention Detection

Naming patterns, export style, barrel exports, path aliases — the tribal knowledge no README captures.

EXPORTS: named (75:18 ratio)
IMPORTS: path alias @/ preferred
BARREL: 40 index.ts files

ANALYSIS_COMPARISON

WHAT_OTHERS_GENERATE

📁 src

📁 components

📄 button.tsx

📄 input.tsx

📁 utils

📄 helpers.ts

"This is a React project using Tailwind CSS."

WHAT_SOURCEBOOK_GENERATES [PRECISION_MODE]

CRITICAL_HUB: lib/dispatcher.ts

Imported by 42 files. Modification risk: HIGH.

CO-CHANGE_COUPLING

Changing auth/provider.ts usually requires updates in middleware/session.ts (88% correlation).

CIRCULAR_DEPENDENCY

models/user.ts → services/auth.ts → models/user.ts

terminal Context injected. Agent ready.

RESEARCH_FOUNDATION

ETH ZURICH, 2025

Auto-generated obvious context makes agents worse

LLM-generated context files reduced task success by 2-3% and increased inference costs by 20%+. Only non-discoverable information improves performance.

KARPATHY, 2026

program.md is the #1 lever for agent effectiveness

Autoresearch ran 700 experiments in 2 days because the curated context file contained only what the agent couldn't figure out alone.

AIDER

PageRank on import graphs for structural importance

Repo-map technique ranks files by how many other files depend on them. sourcebook uses this to identify architectural hubs.

CHROMA RESEARCH, 2026

LLMs lose 30%+ accuracy in the middle of long contexts

sourcebook places critical constraints at the top and bottom of output files — where LLMs pay the most attention.

READY_TO_INDEX?

npx sourcebook init content_copy

VIEW_SOURCE

No API keys. No LLM. Everything runs locally. MIT licensed.