literature-reviewlisted
Install: claude install-skill yushiran/My-Literature-Review-Skills
# Literature review
Two jobs. **Build** a library for a topic, and **answer** questions from a
library that exists. Read [references/workflow.md](references/workflow.md)
once; it is the data contract every script and agent follows.
## Rules that do not bend
1. **Scripts do the deterministic work.** Searching, ranking, downloading,
converting and indexing are `scripts/*.py`. Do not reimplement any of them
in a shell loop or by reading pages yourself.
2. **Agents only where a judgement is needed.** `scout` (sonnet) turns the
topic into research questions and picks the papers; `librarian` (opus)
writes the reading guide. Nothing else calls a model.
3. **Never traverse the library.** A question about the references starts at
`references/<topic>/INDEX.md`. Grep it, pick ids, open only those
`md/<id>/<id>.md`. `INDEX.md` lists the papers that were selected but never
read; do not cite their content as verified.
4. **MinerU needs a token, and the token is the user's to give.** If
`convert.py` exits 3, stop and tell the user:
> MinerU 需要 API token。在 https://mineru.net/apiManage/token 生成一个,然后
> `mineru-open-api auth`(交互式)或 `export MINERU_TOKEN=...`,我再继续。
Do not fall back to `flash-extract` on your own: it drops the figures and
caps at 20 pages. Only if the user says they will not get a token.
5. **Paywalled papers are reachable, and the credentials are the user's to
give.** Read [references/institutional-access.md](references/institutional-acce