ResearchStudio: Microsoft's Claude Code and Codex skills for the research lifecycle

What ResearchStudio is for, and who should care

ResearchStudio is a collection of skills, not a single application. The README describes them as covering the entire research lifecycle, from an under-specified research direction to a published paper, and splits that span into two halves. ResearchStudio-Idea handles the pre-paper half: turning a vague direction into what the README calls a reviewer-defensible, implementable idea, grounded in an empirical taxonomy induced from a corpus of ICLR, ICML and NeurIPS submissions. ResearchStudio-Reel handles the post-paper half: taking a finished paper PDF and producing a print-ready poster, a narrated walkthrough video, a bilingual blog post and an interactive reel viewer.

The intended user is a researcher who already runs an agentic coding tool. The README states the skills run on Claude Code and Codex, so the audience is people who have one of those installed and are willing to hand it a research workflow rather than a codebase. If you want a Python package you import from a notebook, this is the wrong shape of project. If you want a set of prompts and scripts your existing agent can execute, it fits.

How the Idea and Reel halves are structured

The repository is organised as two top-level directories, ResearchStudio-Idea and ResearchStudio-Reel, plus bin, docs, install.sh and package.json. Each half has its own arXiv paper, its own demo page, and its own usage section inside its subdirectory README. That layout matters: the root README is an index, and the operational detail lives one level down.

The Idea half is described in the changelog entry for 2026-08-13. The user query is persisted verbatim and read by every phase that reasons about intent. Papers named in the query without a link are resolved and deep-read. A stated solution direction reaches idea selection, which must record whether it followed or departed from that direction, and why. That is a concrete data-flow choice: intent travels with the pipeline instead of being summarised away at the first step.

The Reel half is described as the post-paper stage and lists four artifacts from one PDF: poster, narrated video, bilingual blog post, interactive reel. A separate release, pptx2video, is offered as a skill or as a CLI that runs without an LLM, which is the one component the README explicitly says can be used without a model in the loop.

Installing ResearchStudio and running a first pass

The README recommends an isolated environment first, through conda or uv. The conda path creates an environment named researchstudio with python>=3.10.

conda create -n researchstudio "python>=3.10" -y
conda activate researchstudio

The uv path creates a virtual environment at .venv and activates it; the README notes the Windows PowerShell activation command as a comment in the same block.

uv venv --python 3.10
source .venv/bin/activate
# For Windows PowerShell:
# .venv\Scripts\Activate.ps1

After the environment, there are two install paths. The first clones the repository and runs install.sh, which the README says installs native tools, Python dependencies, symlinks the skills from the local clone and scaffolds .env.

git clone https://github.com/microsoft/ResearchStudio.git && cd ResearchStudio
bash install.sh

The second is an interactive npx installer that lets you choose which plugins to install, prompts for API keys and writes .env.

npx github:microsoft/ResearchStudio

The package.json confirms the entry point: the bin field maps researchstudio to bin/install.mjs, and the engines field requires node >=16.7.0. After install, the README sends you to the Idea and Reel subdirectory usage sections rather than documenting the invocation at the root. It also recommends model versions at or above claude-opus-4.6 or gpt-5.5. Expect the first run to be a skill invocation inside Claude Code or Codex, not a command in this repository.

Where ResearchStudio stops being the right tool

The root README does not document rollback. install.sh symlinks skills from a local clone and scaffolds .env, but there is no described uninstall path, and the npx installer writes .env as a side effect of an interactive session. If your team needs a reversible install with a documented teardown, that is a gap you will have to work around yourself.

The second limitation is the model dependency. The README states the skills run on Claude Code and Codex and recommends specific model versions. There is no described mode where the Idea or Reel pipelines run without one of those agents. The one exception is pptx2video, which the release note says can run as a CLI without an LLM. So for the core suites, an environment without Claude Code or Codex access is not a supported configuration.

The third is documentation depth. The root README is a pointer document. Usage lives in ResearchStudio-Idea and ResearchStudio-Reel, and the root does not restate it. If you are evaluating this from the repository landing page alone, you will not learn the flags or the expected inputs for either suite. That is a deliberate split, but it makes a shallow read misleading about how much is actually specified.

How it compares with a general-purpose agent workflow

The obvious alternative is doing the same work with a general agent and no skills: paste your draft into Claude Code or Codex, ask for a poster, ask for a blog post, iterate. The difference is grounding. ResearchStudio-Idea is described as grounded in an empirical taxonomy induced from a large-scale corpus of ICLR, ICML and NeurIPS submissions, and its 2026-08-13 release added persistence of the user query verbatim across every phase, resolution of unlinked paper names, and a requirement that idea selection record whether it followed or departed from a stated solution direction.

A general agent has none of that structure by default. It will summarise your intent somewhere in the middle of a long context and lose the constraint you cared about. Whether the taxonomy actually improves the ideas is something the repository's own paper, arXiv 2607.04439, is the place to check, not the README. What the README does establish is that the pipeline keeps intent and provenance as first-class state, which is a different design from a one-shot prompt.

Maintenance, licence and what an upgrade costs you

The repository is not archived, and the last push was on 2026-09-15, which is recent. There are no retrieved releases, so versioning is not visible from the repository listing; the only version string present is 1.0.0 in package.json. Upgrades therefore run through git pull plus install.sh, or through re-running the npx installer, and both touch .env. Because the installer scaffolds .env and the npx path prompts for API keys and writes it, treat that file as state you own and back it up before re-running either path.

The licence is MIT, stated in the README badge and in the LICENSE file at the repository root. MIT is permissive and places few conditions on reuse, but the README does not discuss what happens to artifacts you generate, what data leaves your machine when a skill calls a model, or how API costs are attributed. Those are questions for your own review of the code and your provider terms, not something the licence text answers. Nothing here is legal advice.

Editorial conclusion

Adopt ResearchStudio if you already work inside Claude Code or Codex and want the ideation and post-paper stages scripted rather than improvised: install.sh or the npx installer sets up the skills and a .env, and the README points to the Idea and Reel subdirectories for usage. Do not adopt it if you need a standalone Python library, a documented rollback path, or a tool that runs without an LLM, since the README states the skills run on Claude Code and Codex and recommends model versions at or above claude-opus-4.6 or gpt-5.5. Before committing, verify the .env keys the installer writes against your own provider account, and confirm the Python version your environment resolves to, because the README pins python>=3.10 in the conda path.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. microsoft/ResearchStudio on GitHub
4. Project website
5. README