dailypaper-skills turns one prompt into paper discovery, review and notes

One prompt runs the whole chain

The premise is a single request. The README's headline is 'One daily prompt to find relevant papers, get candid reviews, and turn the best ones into research notes', and the example request is as short as it gets:

Recommend today's papers.

That one prompt runs discovery, then selection, then review, then priority paper notes. The README also notes you can start with a paper you already want to read, so the same bundle handles the daily sweep and the one-off read.

The output is three connected kinds of content rather than one list. Daily recommendations tell you what to read, bucketed as must read, worth a look, or skip, with up to 10 per day by default. Priority paper notes work through the research question, the method, the equations, the figures, the experiments and the limitations. Concept notes and indexes let you follow unfamiliar terms and reconnect new methods with previous reading.

The linking is what makes the third category worth having. Recommendations link to paper notes, and paper notes link to concepts, so the output is a graph of your reading rather than a folder of isolated files. The default folder structure shows it:

your-notes/
├── DailyPapers/
│   └── YYYY-MM-DD-paper-recommendations.md
└── 论文笔记/
    ├── category/MethodName.md
    └── _概念/topic/Concept.md

Existing folder names are preserved when you switch languages, and the README says you can configure your own.

The commentary is the point, not the list of papers

What separates this from a paper feed is stated as a core part of the workflow. Each recommendation explains what the method does, why it relates to your work, what you can reuse, and which claims deserve scrutiny.

The last part is where the rules are strictest. Criticism needs evidence, and anything the abstract leaves unclear is marked for a full-text check. That is a reasonable division of labour: a recommendation based only on an abstract is a guess, and a workflow that produces confident critique from abstracts will eventually hand you a confident wrong claim.

The selection logic is also opinionated in a way worth knowing. Recommendations use history to avoid repeats, and they do not pad a sparse list. A day with three good papers produces three, not three plus filler. The default interests cover LLM reinforcement learning, RL infrastructure, world models and JEPA, dexterous manipulation, and physics simulation, but you choose your own, and the README says a typical request names a directory and a set of interests directly.

Must-read papers continue on to full notes, and existing notes are matched and linked rather than duplicated. So the bundle is designed to run repeatedly against the same vault without producing a pile of near-duplicate files.

Seven skills, and the stages can be rerun on their own

The bundle is seven skills, and the table in the README gives the role of each.

daily-papers runs the complete daily workflow, and paper-reader handles a specific paper with full reading, quick summaries and critical analysis of arXiv papers, local PDFs or Zotero items. daily-papers-fetch does the discovery, selection and enrichment from Hugging Face Daily and Trending and from arXiv. daily-papers-review writes the candid reviews, the reading priorities and the recommendation pages. daily-papers-notes generates the priority notes, adds concepts and backfills links. generate-mocs refreshes the paper and concept indexes. github-trending produces daily, weekly and monthly project lists plus simulation topic searches.

The important architectural point is that fetching, review and note generation connect automatically, but each stage can also be rerun separately. That matters in practice. A review that turns out to be wrong does not require re-running the whole discovery sweep, and a link you broke by moving files can be repaired with the index refresh rather than by starting over.

The everyday prompt table is a good guide to the surface. Beyond the daily and catch-up variants, you can ask to read a specific paper by arXiv link or PDF path, to get the core contribution with a quick summary, to critically analyze a paper, to read a paper from your Zotero library, to show weekly GitHub trending, and to refresh the paper and concept indexes after moving notes.

The repository structure supports this split, with a skills/ directory holding one directory per skill, each with a SKILL.en.md, and shared material under skills/_shared/, including the language rules and the user_config.py script.

Installation is a Python installer with per-agent targets

The requirements are stated plainly, and they are not trivial. You need an agent with file, command and web access, plus Python 3.10 or newer, Git and curl. Full PDF reading additionally needs Poppler, installed with brew install poppler on macOS, sudo apt install poppler-utils on Ubuntu and Debian, or pdftotext and pdfimages on the PATH on Windows. Obsidian and Zotero are optional.

The install is a clone and a Python script:

git clone https://github.com/huangkiki/dailypaper-skills.git
cd dailypaper-skills
python3 install.py --agent codex

You replace codex with claude, cursor, copilot, gemini, opencode or openclaw. On Windows you may need py -3 instead of python3. The README says the installer copies all seven skills, both workflow languages, the templates and the shared dependencies together, which is the right call for something whose stages depend on shared files.

Configuration is a second script, and it is careful about existing state:

python3 skills/_shared/user_config.py --init

The README says an existing configuration is preserved. After that you restart or refresh the agent session and send a prompt naming your output directory and your interests, for example asking to save Markdown to a given path, to work on robot learning, VLA, world models and physics simulation, and to recommend at most 10 papers per day.

The upgrade path is a git pull and the installer with --update, then a session refresh. The README also mentions a Web Viewer as an optional way to read the notes in a browser, with setup in Chinese in the usage guide, and an obsidian-templates/ directory in the repository for Obsidian.

Language follows the prompt, and the model allowance is the budget

Two practical points about cost and language are worth isolating.

On language, the behaviour is that English and Chinese follow your prompt. Ask in English and you get English recommendations and notes, or you can request a language explicitly with 'Read this paper, but write in Chinese'. Switching languages preserves existing folders, note names and link targets, which is the part that would otherwise break a vault, since a renamed note breaks every link to it. The rules live in skills/_shared/language.md, and the README is written in both English and Simplified Chinese with separate README.en.md and README.zh-CN.md files.

On cost, the README is explicit: no extra API key is needed by default. Selection, reviews and full reading use your current agent's model allowance. The budget is therefore the agent subscription you already have, and the cost of a run scales with how many papers get full notes, which is why recommendations are capped at ten per day by default.

That design has a consequence worth stating plainly. A full run reads PDFs, extracts figures and writes several Markdown files per paper, so it is a long agent session rather than a quick call. There is no batch mode and no scheduling described in the README, so the daily run is something you trigger, and a missed day is a missed day.

The other dependency is a live web, since discovery goes to Hugging Face Daily and Trending and to arXiv. The repository topics mention research, and the skill list includes github-trending for the open source side of the same habit.

Where this approach is the wrong tool

The first limit is that everything runs through an agent. A recommendation page produced by a language model is a set of claims, and the ones about relevance and reuseability are judgements that can be wrong in ways that are hard to spot, because they read fluently. The workflow mitigates this with the rule that criticism needs evidence and that anything the abstract leaves unclear is marked for a full-text check, but the mitigation is a prompt instruction, not a mechanism. Anyone treating the must-read tier as a curated reading list is reading the output of a model, not a librarian.

The second limit is the agent sandbox. The skills need file, command and web access, and they run pdftotext and pdfimages on PDFs you give them, and they write into your notes directory. That is a lot of trust placed in a bundle of prompt files, and the value of the whole thing depends on the agent actually doing what the skills say.

The third is the dependency stack for what it is. A paper discovery workflow now requires an agent, Python, curl, Git, Poppler, and often Obsidian or Zotero, and it lives inside a repository that has no GitHub releases, so upgrades come from a git pull rather than a versioned artifact. The repository is not archived and the last push was on 2026-10-15, so it is actively worked on, but there is no changelog and no release history to read.

The fourth is the scope of the sources. Discovery covers Hugging Face Daily and Trending and arXiv, and the github-trending skill covers open source projects. Papers outside those channels, paywalled venues, and work published as blog posts or on personal sites are simply not in the pipeline.

Editorial conclusion

dailypaper-skills suits a researcher who already uses a coding agent and wants paper discovery, selection and note taking folded into the same session, with notes that link back to concepts and to past reading. It is the wrong tool if you want a recommendation feed you can read without an agent, and it is the wrong tool if you cannot let an agent run commands and read PDFs on your machine, since full paper reading depends on Poppler and on web access for arXiv and Hugging Face. Before installing, read the skill definitions rather than only the README, since the individual skills are where the actual behaviour lives, and configure your interests through the shared user_config.py before the first run.

Frequently asked questions

Official sources

1. huangkiki/dailypaper-skills on GitHub
2. Issues
3. License: Apache-2.0
4. README