OneResearchClaw

Start Any Research in One Claw. 🦞

From real materials to reports with a fully autonomous & skill-driven researcher.

OneResearchClaw is a multi-format input → research report generation pipeline designed for real-world workflows.
It supports starting from meeting audio/video, documents, tables, PPT files, ZIP mixed-material packages, and arXiv / YouTube / Bilibili links, then progressively completes content grounding, literature research, quality review, skill evolution, and multi-format export.

🌐 中文 README · 🖥️ Project Page · 🎬 Showcase · 🚀 Quick Start · 🧩 Skills Catalog

💬 WeChat · 📕 RedNote · 🤖 Discord

🎬 Showcase

Generated Report Showcase 5 representative cases across meetings, documents, mixed archives, links, and review workflows — showing multi-topic splitting, controllable research depth, review-rewrite improvement, and multi-format delivery.

🔥 News

• [2026/04/21] Skill-Evolve is officially released: OneResearchClaw now supports continuous skill-set evolution based on user feedback, saving personalized preferences as selectable derived versions so the automation pipeline can gradually adapt to the way you work.

• [2026/04/20] One-click cloud link ingestion: Whether the input is an arXiv paper, a YouTube video, or Bilibili content, simply paste the link and the system will automatically download it locally and start the research workflow.

• [2026/04/19] Multi-model Review mechanism: The Review → Rewrite loop supports bounded iterative revision, significantly improving the quality of final deliverables.

• [2026/04/17] Controllable Research depth: Choose from simple / medium / complex. Whether you need a quick first pass or a deeper research run, the depth and resource usage are easier to control.

• [2026/04/16] Multi-topic parallel execution: The system can automatically identify multiple topics in a meeting and generate independent reports. Even dense group meetings can be preserved in a complete and structured way.

• [2026/04/13] OneResearchClaw: From meeting recordings to PDFs, from tables to ZIP packages, real materials are scattered across formats and locations. Normally, you need to manually connect keyword extraction, research, and summarization across workflow stages. OneResearchClaw solves this problem: input files in different formats, then automatically perform structuring, keyword extraction, deep research, content synthesis, and multi-format report export. Reports are now one command away, without manual handoffs.

✨ Core Highlights

🎯 Typical Use Cases

• 🎙️ Turn meeting audio / video into topic-level reports.
• 📄 Integrate papers, documents, tables, PPT files, and other existing materials into one unified research report.
• 🗂️ Automatically unpack ZIP mixed-material packages and synthesize them into a comprehensive report.
• 🔗 Start content retrieval and analysis from arXiv / YouTube / Bilibili links.
• 📚 Supplement grounded content with literature search and related work analysis.
• 🧪 Improve report quality through a review loop.
• 🧬 Use Skill-Evolve to turn personalized preferences into a new skill subset.
• 📦 Export results as Markdown / DOCX / PDF / PPTX / Audio.

🚀 Quick Start

1. Clone the repository

git clone <your-repo-url>
cd OneResearchClaw

2. Install dependencies

pip install -e .

Optional capabilities can be enabled by module:

pip install -e ".[audio]"
pip install -e ".[document]"
pip install -e ".[export]"
pip install -e ".[table]"
pip install -e ".[api]"
pip install -e ".[full]"

3. Run in Cursor

Start from the entry skill, for example:

.cursor/skills/one-report/SKILL.md

A minimal example prompt:

Please use the existing `.cursor/skills/one-report/` skill to generate a full report from one input file.

First read:
- `.cursor/skills/one-report/SKILL.md`

Input:
- input_path: docs/showcase/inputs/case5/HER2-case5.zip
- output_formats: pdf
- research_mode: complex

Search settings:
- search_backend: cursor
- require_open_link: true
- download_opened_literature: true

Optional:
- output_lang: zh
- transcription_language: zh

require_open_link: true
download_opened_literature: true

output_formats: pdf
output_lang: zh

Tip: Pipeline behavior can be controlled through entry parameters.
Add key: value fields to the one-report prompt to control input, research depth, retrieval method, review mode, and final export format.

Parameter	Default	What it does
input_path	Required	Input material path. Supports local files, ZIP mixed-material packages, and local paths produced after cloud-link downloading.
output_formats	Required	Final export formats. Supports md / docx / pdf / pptx / audio, and comma-separated combinations such as md,pdf,pptx.
research_mode	medium	Controls research depth. Supports simple / medium / complex. This is the core parameter affecting the number of opened papers, retrieval coverage, and token cost.
research_requirements	—	Additional research preferences, such as emphasizing technical contributions, engineering risks, benchmarks, limitations, or application scenarios.
search_backend	cursor	Literature retrieval backend. Supports cursor / external / auto.
require_open_link	true	Whether to require opening and reading sources instead of relying only on search result snippets.
download_opened_literature	true	Whether to download opened literature for later verification and paper-note refinement.
transcription_language	en	Language hint for audio / video transcription. It only affects the ASR stage.
output_lang	en	Final report language. For example, zh means exporting a Chinese report.
reviewer_api_config	Not enabled	Optional external reviewer configuration. When provided, every round of the review → rewrite loop uses the specified external model as reviewer.

research_mode is automatically mapped to runtime parameters in config/research_pipeline.env.
The table below shows the current default preset, used to control the literature target range, minimum number of opened papers, and recent-paper requirements. These values are not fixed rules and can be adjusted in the config according to task needs.

Mode	Literature Target Range	MIN_OPENED_PAPERS	OPEN_TOP_K	MIN_RECENT_PAPERS
simple	3–5 papers	3	2	2
medium	6–10 papers	6	3	4
complex	10–15 papers	10	5	6

🌟 Core Capabilities

1. 🧭 Unified Multi-source Input

OneResearchClaw can start a research workflow from many forms of local materials: meeting recordings, paper PDFs, experiment tables, PPT presentation materials, and even ZIP packages containing multiple files. All of them can enter the same automated pipeline.

The system automatically identifies the input type and routes it to the corresponding grounding skill, converting raw materials into structured content usable by the later research, summary, review, and export stages.

2. 🧠 Multi-Topic Meeting Splitting

In real research and project discussions, a single meeting often contains multiple relatively independent topics: different paper directions, experiment progress, troubleshooting, task planning, or follow-up action items.
If everything is merged into one context, the generated report can suffer from unclear topic boundaries, cross-topic interference, and bloated structure.

For such long meetings and multi-issue discussions, OneResearchClaw can automatically identify multiple topics in the meeting and split each topic into an independent grounded unit.
Each grounded unit keeps the corresponding topic context, key information, and evidence boundary, and can separately enter the research, summary, review, and export workflow.

Meeting Input
audio / video / transcript
        │
        ▼
Meeting Grounding
topic detection + segmentation
        │
        ▼
┌────────────┬────────────┬────────────┐
│ Topic A    │ Topic B    │ Topic C    │
└─────┬──────┴─────┬──────┴─────┬──────┘
      ▼            ▼            ▼
Grounded Unit  Grounded Unit  Grounded Unit
      │            │            │
      ▼            ▼            ▼
Report A      Report B      Report C

3. 🎚️ Controllable Research Depth

Different tasks require different levels of research depth.
Some scenarios only need quick background supplementation, some require coverage of core related work, while complex research questions require more systematic literature mapping.

OneResearchClaw provides three research depth modes through research_mode: simple / medium / complex.
Before running, this parameter is mapped to runtime configuration in config/research_pipeline.env, affecting retrieval scope, number of opened sources, recent-paper requirements, and research cost. The specific preset can be found in the Quick Start parameter table and can also be adjusted according to task needs.

During research, the system controls retrieval, source opening, recent-paper requirements, and evidence organization strength according to the current mode.
For the Cursor-native backend, only sources that are actually opened, read, and saved as local evidence count as valid opened literature. This avoids generating research reports based only on snippet-only search results.

If literature downloading is enabled, the system further downloads opened papers and uses the PDF content to refine paper notes.
Therefore, the final lit.md is not just a search-result summary; it combines queries, opened sources, paper notes, and verifiable materials into a research result.

4. 🔎 Query Confirmation & Directed Search

High-quality research depends not only on how deep the search goes, but also on whether the search direction is accurate from the beginning.
In real use, automatically generated queries may be too broad, too narrow, or fail to match the type of literature the user actually needs. For example, the user may need benchmark papers, survey papers, method papers, system papers, or work related to a specific failure mode / limitation.

Therefore, OneResearchClaw adds a query confirmation step before formal literature retrieval: the system first generates candidate queries based on the grounded note, then supports user confirmation, addition, deletion, or modification.
Only confirmed queries enter the later search, source opening, paper notes, and lit.md generation workflow.

The core value of this mechanism is that users can calibrate the research direction before research cost is actually spent, focusing the budget on the literature types that matter.

With query confirmation, users can more directly specify the research focus they need, for example:

Focus more on recent method papers from the last two years
Add benchmark / dataset related queries
Reduce generic survey queries
Supplement failure mode / limitation directions
Focus on a specific application scenario or model architecture

This brings two direct benefits:

• Reduce ineffective retrieval: avoid queries drifting away from user needs, reducing irrelevant search results and unnecessary source opening;
• Lower token and time cost: especially under medium / complex research modes, calibrating queries early can significantly reduce later ineffective reading and organization.

Confirmed queries serve as the execution basis for the later research stage.
Therefore, OneResearchClaw’s literature research is a directed research workflow that can be manually calibrated before retrieval.

5. 🧪 Multi-model Review Mechanism

Automatically generated research reports often still need further checking:
whether the material coverage is complete, whether the evidence is specific, whether the analysis is deep enough, whether conclusions overclaim, and whether the final content is suitable for delivery.
OneResearchClaw adds a review → rewrite quality loop after summary, enabling structured diagnosis and bounded revision before export.

This stage treats the report as an auditable object and checks it against grounded.md, lit.md, summary.md, and optional paper notes / download manifest for evidence-based quality review.

The review stage supports two reviewer modes:

Revision is performed by an independent writer role. The reviewer is responsible for scoring, diagnosis, and repair actions; the writer rewrites the report according to the repair actions specified by the reviewer.
This role separation makes it easier to connect different models for cross-review and ensures each revision has a clear diagnostic source and repair basis.

Each review round is recorded, making it easier to inspect how the report was modified and how it passed the quality gate. Once the score reaches >= 90, the final report can be exported:

data/review_outputs/<ground_id>/
├── round_0/
│   ├── review_report.md
│   └── review_state.json
├── round_1/
│   ├── review_report.md
│   └── review_state.json
├── review_history.json
├── review_report.md
└── review_state.json

data/reports/<ground_id>/
└── research_report.md

6. ☁️ One-click Cloud Link Ingestion

Besides local files, OneResearchClaw can also start report generation directly from remote links.
When input_path is a URL, the system first calls the remote-input skill to download or parse the remote material into a local file, then hands it to input-router so it can enter the later grounding pipeline according to file type.

This means users can start directly from an arXiv paper, YouTube video, or Bilibili video, without manually downloading, renaming, moving files, or re-specifying downstream processing. Remote materials are first standardized as local inputs, so the later stages still reuse the same grounding → research → summary → review → export workflow.

7. 📦 Multi-format Export

OneResearchClaw supports exporting the reviewed final report into multiple deliverable forms, covering editing, archiving, presentation, and audio-browsing scenarios.
The export stage generates different formats from the same final report, avoiding separate regeneration for each format and preserving version consistency.

Currently supported export formats include:

md / docx / pdf / pptx / audio

Among them, md is suitable for continued editing and version control, docx for formal document delivery and manual editing, pdf for stable reading and archiving, pptx for presentation, and audio for speech-style delivery.

OneResearchClaw also supports controlling the final delivery language through output_lang:

output_lang: zh
output_lang: en

In other words, the same pipeline can export the final report in either Chinese or English as needed.

8. 🧬 Skill-Evolve

The main workflow of OneResearchClaw is responsible for generating reports, while Skill-Evolve helps the workflow gradually adapt to user preferences during use.

In real use, different users often have different preferences for report style, research depth, structure, citation style, and export behavior.
If these preferences are repeatedly specified through temporary prompts, they are difficult to preserve and can easily be lost in later tasks. The goal of Skill-Evolve is to turn recurring feedback into reviewable, verifiable, and reusable skill modification proposals.

It does not directly overwrite the stable main skill set. Instead, it first generates a patch proposal based on user feedback, then produces a derived skill subset after validation. This preserves the stability of the original workflow while allowing users to maintain customized versions based on their own usage habits.

user feedback → patch proposal → validation gate → derived skill set

Skill-Evolve mainly supports:

• Analyzing which skill needs adjustment based on user feedback, such as insufficient report depth, mismatched structure, review dimension changes, or export improvements;
• Generating patch proposals for specific skills, rather than only producing a temporary prompt;
• Using regression / validation gates to check whether modifications break the original workflow;
• Producing derived skill subsets that can be selected for future tasks;
• Persisting personalized preferences at the skill level rather than leaving them in a single conversation.

🧩 Skills Catalog

The table below lists the main skill set in the current repository, grouped by function.

⭐ Star History

🙏 Acknowledgements

OneResearchClaw is inspired by several excellent open-source projects and skill-based workflow practices:

• AutoResearchClaw
• Auto-claude-code-research-in-sleep
• A-Evolve
• awesome-claude-skills

We thank the authors and maintainers of these projects for sharing their work with the open-source community.

📄 License

MIT — see LICENSE for details.

📌 Citation

If you find OneResearchClaw helpful, consider cite us:

@misc{OneResearchClaw2026,
  author       = {Yang, Wei* and Zhao, Yiming* and Zeng, Yu* and Fang, Zhen* and Huang, Wenxuan and Zhang, Ziao and Shi, Kou and Miao, Qing and Cai, Jiayin and Zhao, Jiawei and Bao, Xikun and Chen, Lin and Chen, Zehui and Feng Zhao},
  title        = {OneResearchClaw: Fully Autonomous, Skill-driven Research Synthesis from Heterogeneous Inputs},
  year         = {2026},
  organization = {GitHub},
  url          = {https://github.com/gaotiexinqu/OneResearchClaw},
}

_{Built with Cursor by the OneResearchClaw Team}