Junshi skill
The research-junshi skill for Claude Code and Codex tracks daily papers for your topics and proposes ranked research ideas.
- 134 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The skill reads your papers and keeps an interest profile in `~/.junshi`.
- `setup_automation.sh` adds a cron job after confirmation.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/junshiDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The skill reads your papers and keeps an interest profile in `~/.junshi`.
- `setup_automation.sh` adds a cron job after confirmation.
This is third-party code. Review the repository files before installing.
What it does
Junshi reads your papers, remembers research interests and feedback and tracks new literature. From that it builds three to five ideas ranked by score: each has a pitch, cited motivation, a link to your work, a first experiment and the main threat to success. The score adds novelty at weight 0.4, feasibility at 0.3 and impact at 0.3. Memory keeps paper metadata, version history, DOI and arXiv identifiers, past recommendations and rejected directions in `JUNSHI_HOME` (default `~/.junshi`). A dependency-free Python collector pulls arXiv and Crossref and writes a digest to `~/.junshi/digests/`; it runs without a model or agent permissions. Ideas are produced by an interactive agent on request.
Who it is for. Researchers and PhD students who want a daily paper selection for their topics and draft directions with a first experiment and an assessment of weak spots.
Good fit when
- You want a morning digest of new arXiv and journal papers for your interests.
- You need to remember which directions were rejected and why.
- After a finished project you want a next step that respects your resources.
- One researcher memory is shared by Claude Code and Codex.
Not a fit when
- You need a finished literature review with every citation verified: the skill selects and ranks, it does not replace reading.
- Your field is outside arXiv and Crossref: sources are limited to those two.
- Cron cannot run on your work machine.
Example request
Run research-junshi. I work on causal inference, my papers are in ~/papers/, propose three directions for today.Limitations
Python 3.9 or newer is required; PDF text extraction needs poppler installed separately. Conference coverage is not automatic and sources have explicit caps. The daily cron runs only while the machine is on and online. Claude Code and Codex installs share `JUNSHI_HOME`, so keep both on the same repository revision. The last commit is from 16 September 2026. Proposed ideas rest on retrieved metadata and abstracts, so verify citations and novelty against the original sources.
How to disable. Remove the `research-junshi` folder from `~/.claude/skills/` or `~/.agents/skills/` and the cron line added by `setup_automation.sh`. Delete the `~/.junshi` memory directory separately.
FAQ
How do I enable the daily digest?
Run `python3 scripts/daily.py` to verify sources, then `bash setup_automation.sh`. The script previews the cron entry and asks for confirmation.
Are API keys needed?
No. The collector uses arXiv and Crossref with no model or credentials; ideas come from your regular agent.
Where are digests and memory?
In `~/.junshi`: digests in `digests/YYYY-MM-DD.md`, and the path changes through `JUNSHI_HOME`.
How do I reject a direction?
Tell the agent, for example "Reject large language model training: it exceeds my compute budget". The reason is stored and shapes later ideas.
How do I connect it to Codex?
Clone the repository into `$HOME/.agents/skills/research-junshi` and invoke `$research-junshi`. If the skill does not appear, restart the session.
Discussion
No questions or reviews yet.
Related
SkillsA plugin for academic work: dozens of skills for literature review, experiments, data analysis and paper writing
A skill picks five recent papers and one classic weekly, learns from your choices and writes a bilingual Feishu deep read.
Twenty-seven Chinese-language skills for development, thinking, learning, writing, visuals, branding and working with agents
Picks fresh arXiv and HuggingFace papers each day, writes a critical review and saves Markdown notes with formulas.
The author's personal skills for reading papers and books, dissecting concepts, writing and PNG cards; instructions in Chinese, installed via skills CLI
Deep Research Skills
Deep Research Skill
A structured deep-research skill for Claude Code, OpenCode and Codex with human control at every step
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Junshi (this card) | no | 134 | |
| dailypaper-skills | no | 1,356 | skills for daily discovery and critical reviews; Junshi adds memory of interests and rejections and ranked ideas |
| flonat-research | no | 151 | infrastructure of skills, hooks and rules for PhD researchers; Junshi focuses on literature selection and ideas |
README in short
Features, Claude Code and Codex install, request examples, idea scoring, daily automation, migration from the old store and tests.