From real materials to reports with a fully autonomous & skill-driven researcher.
OneResearchClaw is a multi-format input β 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.
π δΈζ README Β· π₯οΈ Project Page Β· π¬ Showcase Β· π Quick Start Β· π§© Skills Catalog
π¬ WeChat Β· π RedNote Β· π€ Discord
-
[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 β 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 Capability | Description |
|---|---|
| π§ Unified Multi-source Input | Supports many types of materials, from meeting recordings to PDFs, from tables to ZIP packages. The system automatically identifies the material type and routes it to the corresponding grounding skill, reducing manual organization, conversion, and preprocessing costs. |
| π§ Multi-Topic Meeting Splitting | For long meetings, complex discussions, or multi-issue materials, the system can automatically split content into multiple topics and generate independent research branches. Each topic can separately go through grounding, research, summary, and export, making it suitable for producing multiple focused reports from one complex meeting. |
| ποΈ Controllable Research Depth | Supports simple / medium / complex research modes to control literature search, source opening, evidence organization, and report analysis depth. It can generate a quick briefing or a more complete research report. |
| π Query Confirmation & Directed Search | Before formal retrieval, the system first generates candidate queries based on the grounded note and supports user confirmation, addition, deletion, or modification. This makes literature search better aligned with the user's real research direction and reduces irrelevant retrieval and token waste. |
| π§ͺ Multi-model Review Mechanism | After report generation, the report enters a review β rewrite loop. Different models or review perspectives can check completeness, evidence consistency, logical coherence, and over-claiming, then revise the report through bounded rounds. |
| βοΈ One-click Cloud Link Ingestion | Supports cloud link ingestion and processing for sources such as arXiv, YouTube, and Bilibili, allowing online materials to enter the same research pipeline directly. |
| π¦ Multi-format Export | Supports exporting final results as md / docx / pdf / pptx / audio, covering reading, archiving, presentation, and reuse scenarios, with both Chinese and English export supported. |
| 𧬠Skill-Evolve | Supports evolving skill subsets based on user feedback through feedback collection, patch generation, regression checks, and version promotion, so the system can continue improving rather than staying fixed as a one-off hand-written workflow. |
- ποΈ Turn meeting audio / video into topic-level reports.
- π Integrate papers, documents, tables, PPT files, and other existing materials into one unified research report.
- ποΈ Automatically unpack ZIP mixed-material packages and synthesize them into a comprehensive report.
- π Start content retrieval and analysis from arXiv / YouTube / Bilibili links.
- π Supplement grounded content with literature search and related work analysis.
- π§ͺ Improve report quality through a review loop.
- 𧬠Use Skill-Evolve to turn personalized preferences into a new skill subset.
- π¦ Export results as Markdown / DOCX / PDF / PPTX / Audio.
git clone <your-repo-url>
cd OneResearchClawpip 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]"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
π Expand: Case 5 full execution path and key intermediate artifacts
Below, Showcase Case 5 Β· Mixed Materials β Bilingual Final Delivery is used as an example to show the main stages of a one-report pipeline run, from a ZIP mixed-material input to the final PDF.
For easier verification, key intermediate artifacts are placed under docs/showcase/reports/readme_case/. The file names mentioned below can be clicked to view the corresponding artifacts.
| Config Item | Current Value |
|---|---|
| Input material | docs/showcase/inputs/case5/HER2-case5.zip |
| Input type | ZIP mixed-material package |
| Research mode | complex |
| Search backend | cursor |
| Require opened sources | require_open_link: true |
| Download opened literature | download_opened_literature: true |
| Output language | zh |
| Output format | pdf |
| Showcase ground id | archive-HER2-case5_20260415130231 |
one-report first reads input_path, detects that the input is a ZIP mixed-material package, and routes the task to the archive grounding workflow.
This stage unpacks the material package, enumerates internal files, and records which grounding skill each file is routed to.
You can inspect manifest.json and routed_items.json to see how the input materials are identified and routed.
The grounding stage converts different files inside the ZIP package into structured content that can be used by later research, summary, and review stages.
The most important artifact is grounded.md.
It summarizes the input material's themes, key information, researchable questions, and context needed for report generation.
To further inspect the processing result of each individual file inside the ZIP package, open the child_outputs/ directory.
The README does not need to expand these child outputs one by one; they mainly show that different files inside the mixed-material package are grounded separately and then aggregated into the main grounded.md.
The research stage does not start searching immediately. Instead, it first generates candidate queries based on grounded.md.
Users can confirm, add, delete, or modify queries before formal retrieval, reducing irrelevant search and token waste.
Two files are worth checking at this stage:
queries.json: candidate queries generated from the grounded note;queries_confirmed.json: the confirmed queries that actually enter the retrieval stage.
This step reflects OneResearchClaw's directed research capability: calibrate the search direction first, then spend retrieval, source-opening, and reading costs.
Under complex research mode, the system performs a deeper literature research process.
Because the current settings are:
require_open_link: true
download_opened_literature: true
the system does not rely only on search snippets. It tries to open sources, preserve evidence, and download opened literature for later verification.
At this stage, you can inspect:
search_results.json: records retrieval results;opened_sources/: stores the contents of sources that were actually opened;opened_paper_notes.jsonl: aggregates paper-level notes;lit_initial.md: the initial literature research result;manifest.json: records literature download status.
Finally, the research stage generates lit.md.
This is the most important external literature basis for the later summary stage, combining queries, opened sources, paper notes, and research conclusions.
The summary stage reads grounded.md and lit.md, then generates the first report draft.
The goal of this stage is not to compress the materials into a few bullet points, but to integrate the original materials and literature research results into a structured research report.
The corresponding artifact is summary.md.
After the initial report is generated, it can enter the review β rewrite loop.
The reviewer checks topic alignment, coverage, evidence specificity, analytical depth, structure coherence, and deliverability. The writer then revises the report according to the reviewer's repair actions.
At this stage, you can inspect:
review_history.json: records multi-round review / rewrite history;review_report.md: the final review report;review_state.json: the final review state;round_0/,round_1/,round_2/: detailed records for each review round.
When the review reaches the passing condition, the system generates the final report: research_report.md.
All downstream export formats are based on this reviewed final report, not on the initial summary draft.
The export stage reads research_report.md and exports the requested format according to output_formats and output_lang.
The current example is configured as:
output_formats: pdf
output_lang: zh
So it generates a Chinese PDF report: zh/report.pdf.
An English version is also kept in the showcase: en/report.pdf.
Tip: Pipeline behavior can be controlled through entry parameters.
Addkey: valuefields to theone-reportprompt to control input, research depth, retrieval method, review mode, and final export format.
Parameter Default What it does input_pathRequired Input material path. Supports local files, ZIP mixed-material packages, and local paths produced after cloud-link downloading. output_formatsRequired Final export formats. Supports md / docx / pdf / pptx / audio, and comma-separated combinations such asmd,pdf,pptx.research_modemediumControls research depth. Supports simple / medium / complex. This is the core parameter affecting the number of opened papers, retrieval coverage, and token cost.research_requirementsβ Additional research preferences, such as emphasizing technical contributions, engineering risks, benchmarks, limitations, or application scenarios. search_backendcursorLiterature retrieval backend. Supports cursor / external / auto.require_open_linktrueWhether to require opening and reading sources instead of relying only on search result snippets. download_opened_literaturetrueWhether to download opened literature for later verification and paper-note refinement. transcription_languageenLanguage hint for audio / video transcription. It only affects the ASR stage. output_langenFinal report language. For example, zhmeans exporting a Chinese report.reviewer_api_configNot enabled Optional external reviewer configuration. When provided, every round of the review β rewrite loop uses the specified external model as reviewer.
research_modeis automatically mapped to runtime parameters inconfig/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_PAPERSOPEN_TOP_KMIN_RECENT_PAPERSsimple3β5 papers 3 2 2 medium6β10 papers 6 3 4 complex10β15 papers 10 5 6
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.
| Input Category | Supported Formats | Typical Content |
|---|---|---|
| Audio | .mp3, .wav, .m4a |
Meeting recordings, interviews, voice notes |
| Video | .mp4, .mov, .mkv |
Meeting videos, lecture recordings, demo recordings |
| Documents | .pdf, .docx, .md, .txt |
Papers, reports, notes, documentation |
| Tables | .xlsx, .csv |
Data sheets, experimental results, statistical tables |
| Presentations | .pptx |
Presentation materials, course slides, project decks |
| Archives | .zip |
Multi-file mixed-material packages |
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
β
βΌ
Meeting Grounding
topic detection + segmentation
β
βΌ
ββββββββββββββ¬βββββββββββββ¬βββββββββββββ
β Topic A β Topic B β Topic C β
βββββββ¬βββββββ΄ββββββ¬βββββββ΄ββββββ¬βββββββ
βΌ βΌ βΌ
Grounded Unit Grounded Unit Grounded Unit
β β β
βΌ βΌ βΌ
Report A Report B Report C
| Application Need | Processing Method |
|---|---|
| One meeting contains multiple research directions | Automatically identify topics and split them into multiple grounded units |
| Different topics require independent research | Each topic can separately enter the research stage and generate its own literature result |
| Group meetings / interviews / discussions need multiple outputs | One meeting can produce multiple focused reports by topic |
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.
| Research Mode | Suitable Scenario |
|---|---|
simple |
Quickly understand background and supplement a small amount of core related work; suitable for lightweight briefing or initial exploration. |
medium |
Balance coverage, cost, and report depth; suitable as the default research mode. |
complex |
Suitable for cross-direction, complex problems, or tasks requiring systematic literature mapping; useful for generating deeper research reports. |
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.
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.
| Query Type | Main Purpose |
|---|---|
problem_queries |
Supplement problem background, domain context, benchmarks, and task definitions. |
method_queries |
Retrieve methods, baselines, solutions, and related technical routes. |
constraint_queries |
Retrieve risks, constraints, failure modes, data-quality issues, and open problems. |
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 / complexresearch 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.
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 β 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.
| Review Dimension | What it Checks |
|---|---|
topic_alignment |
Whether the report stays aligned with the grounded topic and original material requirements. |
coverage_completeness |
Whether it covers important themes, key papers, and main issues in lit.md. |
evidence_specificity |
Whether key conclusions are supported by grounded notes, literature results, or opened evidence. |
analytical_depth |
Whether the report explains why the problem matters, how far the evidence supports the claims, and what uncertainties remain. |
structure_and_narrative_coherence |
Whether the report structure is clear, and whether there are repetitions, jumps, or mixed topics. |
deliverability |
Whether the report is ready for later md / docx / pdf / pptx / audio export. |
The review stage supports two reviewer modes:
| Reviewer Mode | Description |
|---|---|
| Cursor local reviewer | Uses the reviewer subagent defined in .cursor/agents/reviewer.md to review the report locally. |
| External reviewer API | When reviewer_api_config is provided, external models such as OpenAI or Gemini can be used as reviewers, with the same external reviewer backend maintained across each review round. |
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>/
βββ round_0/
β βββ review_report.md
β βββ review_state.json
βββ round_1/
β βββ review_report.md
β βββ review_state.json
βββ review_history.json
βββ review_report.md
βββ review_state.json
data/reports/<ground_id>/
βββ research_report.md
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 β research β summary β review β export workflow.
| Remote Source | Supported Form | Standardized Result |
|---|---|---|
| π arXiv | arxiv.org/abs/..., arxiv.org/pdf/... |
Downloaded as a PDF file |
| π¬ YouTube | youtube.com/watch, youtu.be, youtube.com/shorts |
Downloaded as a video file |
| π¬ Bilibili | bilibili.com/video/BV..., av..., b23.tv/... |
Downloaded as a video file |
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.
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 β patch proposal β validation gate β 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.
The table below lists the main skill set in the current repository, grouped by function.
| Skill | Description |
|---|---|
one-report |
Main entry skill that organizes the full report generation pipeline |
input-router |
Identifies input type and routes the task to the appropriate skill |
remote-input |
Handles remote link inputs such as arXiv / YouTube / Bilibili |
archive-grounding |
Handles ZIP / archive mixed-material packages and aggregates multiple internal contents |
document-grounding |
Handles PDF / DOCX / Markdown / TXT document inputs |
table-grounding |
Handles XLSX / CSV table inputs |
pptx-grounding |
Handles presentation inputs |
meeting-grounding |
Main grounding entry for meeting-type materials |
meeting-audio-grounding |
Audio meeting grounding |
meeting-video-grounding |
Video meeting grounding |
audio_structuring |
Audio content structuring, transcription, and segmentation capabilities |
grounded-research-lit |
Runs research based on grounded notes, opens sources, and generates lit.md |
grounded-summary |
Integrates grounded content and research evidence to generate a comprehensive report |
grounded-review |
Runs the review β rewrite loop to improve report quality |
report-export |
Exports final formats such as md / docx / pdf / pptx / audio |
skill-evolve |
Generates patch proposals based on user feedback and builds new skill subsets |
OneResearchClaw is inspired by several excellent open-source projects and skill-based workflow practices:
We thank the authors and maintainers of these projects for sharing their work with the open-source community.
MIT β see LICENSE for details.
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


