Research Junshi — Development skill for Claude Code

🪭 Junshi (军师)

**Your personalized research strategist**

*Not just what's new. What's new for you.*

Junshi is a Claude Code skill for researchers. It reads your papers, builds a profile of your methods and interests, tracks new papers from arXiv and the venues you care about, and proposes 3–5 ranked research directions you can actually test.

The product is **Junshi (军师)**. The Claude Code skill name is **`research-junshi`**.

You stay the researcher. Junshi acts like a daily 军师: it connects your past work to fresh literature, looks for gaps, and turns them into concrete ideas with a first experiment and a main risk.

It works across many fields, including machine learning, statistics, economics, biology, physics, robotics, and more.

30-second example

Tell Claude Code your situation in plain language:

I work on causal inference and econometrics. My papers are in ~/papers/.
I'm thinking about better ways to handle high-dimensional confounders.
Run research-junshi.

Claude reads your papers, builds a profile, searches today's arXiv and recent venue papers, and saves a digest like this:

### [Rank 1] Debiased Lasso Meets Synthetic Control
Score: Novelty 4/5 · Feasibility 5/5 · Impact 4/5 → 4.3/5

The pitch: Synthetic control methods break down when the donor pool is large
relative to the pre-treatment window. Your debiased Lasso work already handles
high-dimensional nuisance estimation — applying it to reweight the donor pool
gives a synthetic control estimator that is valid even when p >> T.

Why now: Arkhangelsky et al. (2021) opened the synthetic DiD direction but left
the high-dimensional donor case open. Two AER papers this month circle the same
gap from different angles.

First experiment: Simulate a panel with n=500 donors, T=50 periods, and sparse
true weights. Compare your debiased estimator against standard synthetic control
and SDID on coverage and RMSE. One afternoon of code.

Main risk: The weights may not sum to 1 after debiasin