Junshi — Daily Research Strategist
Turns Claude Code into a personal research strategist: it reads your papers, watches arXiv daily, and delivers ranked, testable research ideas.
Data flow is transparently disclosed (writes only to ~/.claude/research-junshi/, no upload/telemetry); setup script checks and confirms replacement of existing cron jobs; both SKILL.md and README explicitly warn that --dangerously-skip-permissions bypasses all permission prompts with no technical scope enforcement. Deducted because the automation path genuinely relies on unscoped headless execution, isolation/recovery depends on user discipline, and the unverified publisher makes attribution and maintenance responsibility uncertain.
Instructions are self-consistent; setup_automation.sh has clear failure branches (missing claude CLI, missing profile) with intelligible messages and duplicate-cron detection. Deducted under the static 10-cap: external dependencies (arXiv API, WebSearch/WebFetch, poppler) have no test coverage; references/venues.md is cited but absent from the evidence, so a key referenced file cannot be confirmed; abnormal-input handling (bad PDFs, empty searches) is left to the model.
Trigger conditions in the frontmatter are specific (research ideas, literature tracking, digests); domains and output format are clear. Deducted because core function depends on network reachability of arXiv API and web search, which may be unstable for mainland-China users; boundary-condition evidence is limited.
Well-layered docs (README + SKILL.md), full Apache-2.0 license, clear install and dependency notes, generated-file locations explained. Deducted for missing versioning, changelog, and stated maintenance/update path, and for the referenced references/venues.md not appearing in the evidence (hidden assumption).
The workflow is complete and the digest template is directly usable (scores, first experiment, risk); README shows an example output. Deducted under the static 7-cap: core value (idea quality, paper relevance) depends entirely on daily LLM execution with no verified representative output, and comparative benefit over manual literature review is unevidenced.
Auditable source code, script, and an example output exist. Deducted under the static 5-cap: no tests, no CI, no third-party execution evidence; the sample digest is illustrative, not a reproducible result.
- Automation mode uses --dangerously-skip-permissions, bypassing all permission prompts with no technical scope enforcement; review setup_automation.sh before enabling and use only in a trusted local environment.
- The referenced references/venues.md was not present in the reviewed evidence; verify it exists after installation.
- Core function depends on the arXiv API and web search, which may be unreliable for mainland-China users.
- Daily generated 'research ideas' are unverified; treat them as brainstorming input only, not as research decisions.
- Publisher identity is unverified, with no versioning or stated update path; consider pinning a specific commit.
What does this skill do, and when should you use it?
Junshi (军师, "strategist") is a Claude Code skill for researchers in any field. On first run it reads your paper PDFs and builds a research profile covering your methods, open problems, and preliminary results. Each day it then pulls fresh arXiv papers via the arXiv API plus recent papers from your target venues via web search, summarizes the most relevant ones, and generates 8–10 research ideas grounded in your profile. Ideas are scored on novelty, feasibility, and impact, and the top 3–5 are written into a daily digest. It is not a literature-summary tool — it behaves like a blunt senior collaborator pushing bold, testable directions.
On setup, it conversationally collects your research area, target venues, arXiv categories, and papers folder, parses your PDFs with the Read tool or pdftotext, and saves a profile and config under ~/.claude/research-junshi/. Each daily run: queries the arXiv API with WebFetch (two searches, ~100 candidates, 10 selected); runs parallel WebSearch queries against your target venues using patterns from references/venues.md (3–5 papers selected); summarizes each paper's core idea, key insight, open questions, and relevance; generates 8–10 ideas informed especially by your logged preliminary results; scores them as Novelty×0.4 + Feasibility×0.3 + Impact×0.3; and saves the digest to ~/.claude/research-junshi/digests/YYYY-MM-DD.md. An optional setup_automation.sh installs a cron job for fully unattended daily runs.
- A researcher switching fields: point the skill at a folder of papers in the new area to quickly build a profile and get entry-point ideas.
- A grad student tracking the literature: receive a morning digest of what's new in their area (e.g., causal inference, NLP) and what actually matters for their problem.
- Someone with unpublished preliminary results: hand over surprising observations and get interpretations of what they imply and what's missing to make them publishable.
- A postdoc stuck on topic selection: get cross-paper, gap-exploiting ideas each with a first experiment and a main risk.
- Anyone who doesn't want to manually scan arXiv: set up the cron job for automatic daily digests without opening Claude Code.
- Researchers outside ML (economics, biology, physics, robotics): rely on built-in venue knowledge covering Econometrica, Nature, PRL, ICRA, and more.
What are this skill's strengths and limitations?
- Ideas are grounded in your own papers and preliminary results rather than generic summaries; each ranked idea includes a first experiment and main risk.
- Covers both arXiv (freshness) and peer-reviewed venue papers (depth), adapting to whatever field and venues you configure.
- Transparent scoring (novelty 0.4 / feasibility 0.3 / impact 0.3) that is customizable by editing SKILL.md.
- Offers a fully automatic cron mode so daily runs require no manual triggering.
- The automation mode runs with --dangerously-skip-permissions, bypassing all permission prompts with no technical scope enforcement — the author himself warns to review the script and use only in a trusted local environment.
- Depends on WebFetch/WebSearch and live retrieval (arXiv API, search engines); the source provides no tests or accuracy data on retrieval or idea quality.
- No test suite; idea quality rests entirely on model judgment with no external evaluation evidence.
- The repo lists no topics, so community size and activity are unknown.
How do you install this skill?
1) Clone the repo into the Claude Code skills directory: git clone https://github.com/junshi-research/research-junshi.git ~/.claude/skills/research-junshi
2) Reload plugins in Claude Code with /reload-plugins.
3) For PDF reading, install poppler: brew install poppler (macOS) or apt install poppler-utils (Linux).
4) (Optional) For fully automatic daily runs, execute bash ~/.claude/skills/research-junshi/setup_automation.sh to install a cron job.
How do you use this skill?
On first run, describe your situation in plain language, e.g.: "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 asks a few follow-ups (skippable — it makes confident defaults), reads your papers, builds the profile, and produces today's digest. Daily: say "Give me today's research digest." Update your profile with "Update my Junshi profile...". For unattended runs, run setup_automation.sh and pick a time; digests land in ~/.claude/research-junshi/digests/ each morning.