kw-engine

---

The problem

You read a paper, extract a clever trick, and forget it. Six months later you face a problem the same trick would solve — but it was in another field, used different words, and your notes are a pile of PDFs. Your bottleneck was never finding papers. It was reusing the underlying method across domains.

kw-engine treats that as the actual problem.

What it does

It distills literature through three layers, stripping the domain and keeping the transferable logic:

 Paper PDF
   │
   ├─  L1  faithful extraction   what the paper says, with section locators — no interpretation
   │
   ├─  L2  abstraction           strip the domain, keep the transferable core:
   │                               problem-signature   ·  WHEN it applies (problem structure)
   │                               ↔ mechanism + math  ·  WHAT to do
   │                               ↔ rationale         ·  WHY structure ↔ mechanism holds
   │
   └─  L3  synthesis             cluster principles into a design-space map; surface
                                  contradictions and GAPS — your next reading list

Then the payoff:

New problem  →  search by its structure  →  matched mechanism + rationale + when-it-breaks

It's not a search index over text. It's a compiler from empirical results to reusable problem-solving strategies. (The math of why this works is at the end.)

How it compares

	What it retrieves	Indexed by
RAG / vector search	text chunks	surface semantics
Zotero / citation managers	references & PDFs	metadata, tags
Skill libraries (e.g. Voyager)	executable task code	task name
kw-engine	mechanism + why-it-works + when-it-fails	problem structure

---

Get started

kw-engine runs inside Claude Code. Install it once, then you just talk to Claude about your papers.

Brand-new to all this? Follow the step-by-step beginner tutorial — it assumes zero terminal experience and walks through a full case. Also available in 中文.

Install (one time)

Four lines in your terminal — the only commands you'll ever need:

curl -LsSf https://astral.sh/uv/install.sh | sh              # install uv (skip if you have it)
uv tool install git+https://github.com/chenpg2/kw-engine     # the engine
claude plugins marketplace add chenpg2/kw-engine             # the Claude Code plugin…
claude plugins install kw-engine@kw-engine                   # …skills + agents that drive it

Windows: replace the first line with powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

After this, everything happens in conversation.

No Claude subscription? Use Kimi or DeepSeek. Claude Code can run on any Anthropic-compatible backend. Point it at your provider with two env vars in ~/.claude/settings.json and everything below works identically:

• Kimi: ANTHROPIC_BASE_URL=https://api.moonshot.cn/anthropic, ANTHROPIC_AUTH_TOKEN=<your Kimi key>
• DeepSeek: ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic, ANTHROPIC_AUTH_TOKEN=<your DeepSeek key> (it even maps strong/cheap models, so the read-cheap/distill-strong split is preserved)

Then just talk to Claude

Open Claude Code in any folder and say what you want. Claude uses the engine for you and shows you the results — you never touch a file or a command.