Maiwen
Find worthwhile research questions your dataset can answer.
You have a dataset. What would be worth learning from it, and which questions can its measurements actually resolve?
Maiwen helps develop those questions inside Codex or Claude Code. It studies the scientific field, learns from how strong papers asked their questions, and examines the data and their previous uses. Separate agents propose ideas, investigate promising comparisons and review the resulting research plans. You can start with a broad interest or with no question in mind.
Install · Read an example · Explore the workflow
Research preview · 0.2.0-beta.3. Get started · Validation and limitations
Install with your agent
Open a local Codex or Claude Code session and paste:
Install Maiwen from https://github.com/BeibaiDraco/maiwen
for the host I am using (Codex or Claude Code).
Read QUICKSTART.md and install the current research-preview release.
Verify the installed version and the availability of:
discover, learn-patterns, dataset-card, and prior-use.
Tell me whether I need to reload plugins or start a new session.
If a manual step is needed, walk me through it.
Installation is available directly from GitHub and does not depend on an official directory listing. The quickstart also provides terminal commands and instructions for a downloaded copy.
Start with your data
After installation, open your research workspace and ask:
Use Maiwen to find worthwhile questions this dataset can answer:
/absolute/data/path. I have no seed question. Start by checking the inputs and agreeing on a research budget.
You can also supply local papers, known prior work, a direction or an existing library of question patterns. Maiwen first reads the available material, then asks about missing facts that need your knowledge. You can steer, pause or resume the work in the same conversation.
Read the scientific argument
For each developed question, the final report explains what is unknown, why an answer matters, what earlier studies established, and how the proposed computation would answer it. It describes the measurements, assumptions and possible interpretations in enough detail for a colleague to assess the study.
The report links to the field overview, paper notes, data checks, candidate history and reviews. You can inspect what informed an idea and why it was kept, changed or rejected. All saved research documents are in English.
The IBL example explains a real question about neural timescales: why it matters, how paired recordings could answer it, what different results would mean, and which measurement choices still need work. The measurement walkthrough shows how a broad idea becomes a precise comparison and a proposed calculation.
How Maiwen develops a question
flowchart LR
A[Dataset and optional direction] --> B[Field knowledge]
A --> C[Question patterns from papers]
A --> D[Data and previous studies]
B --> E[Independent review]
C --> E
D --> E
E --> F[Generate a broad set of questions]
F --> G[Screen and investigate]
G --> H[Develop promising questions]
H --> I[Review and revise]
I --> J[Research plans with their evidence]
Ideas can come from experiments, theory, connections across fields or an unusual feature of the data. Planning then asks whether the proposed comparison would teach us something useful and whether the available observations support it. When an idea changes substantially, its scientific contribution and data requirements are checked again.
Review and revision have a finite budget. If the first set cannot advance, Maiwen may explore one further set of ideas informed by those failures. The report can contain supported recommendations, conditional directions or reasons to stop; the number follows the evidence. See the complete workflow and guide to reading results.
Four skills, one workflow
| Skill | What it does |
|---|---|
discover | Coordinates the full process from preparation to reviewed research plans |
learn-patterns | Builds a reusable library of ways to form questions and hypotheses from papers |
dataset-card | Examines what the data measure, how they were collected and which comparisons they support |
prior-use | Explains previous uses of the data and finds the closest work on a proposed question |
A pattern library can travel between projects. Bring your own PDFs or let the agent find instructive studies. For each new dataset, Maiwen combines that library with current field knowledge and direct investigation of the data.
Use your existing research tools
Codex or Claude Code supplies the model, tools, permissions and separate agent sessions. Maiwen supplies the research instructions and document structure. It uses your existing host account; there is no separate model service to configure. As the host's reasoning, search and coding capabilities improve, Maiwen can use them directly.
The workflow develops research plans. Planning includes permitted data inspection and synthetic checks; running the proposed scientific analysis requires a separate request. The validation record explains what has been tried so far and what remains uncertain.
From Maieusis to Maiwen
Maieusis is Maiwen's predecessor. It introduced the central idea of learning how successful papers form questions and developing those questions against real datasets. Maiwen carries that purpose into a plugin that works within Codex and Claude Code. Read Method and lineage for the scientific rationale and design history.