Microsoft has open-sourced ResearchStudio-Reel, a pipeline of five Claude Code and OpenAI Codex skills that turns a finished document into a print-ready poster, a narrated talk video, and a bilingual blog post in one pass. Released on July 5, 2026 alongside its arXiv paper and a public GitHub repository under an MIT license, the system reports a 84 percent to 93 percent win rate against prior automation tools and frontier language models across held-out judges. For creators and builders who spend hours reformatting one piece of source material into slides, video, and social copy, it is the clearest example yet of agent skills wired into a real content-production workflow.
What ResearchStudio-Reel Actually Does
The tool attacks what its authors call the "last mile" of dissemination: the manual grind of turning one dense document into every format an audience expects. Instead of treating a poster, a video, and a blog as three separate jobs, ResearchStudio-Reel extracts the source once and feeds a shared asset layer into every generator. That single design choice is why the outputs stay factually consistent with each other rather than drifting the way three independent tools would.
The project ships as part of a larger suite documented on the ResearchStudio project page. A companion module, ResearchStudio-Idea, handles the "first mile" of turning a rough problem into a structured idea card. Reel is the post-publication half, and it is the piece that matters for anyone whose job is repackaging finished work into shareable media.
The Five-Skill Pipeline, Step by Step
ResearchStudio-Reel is built as five composable skills rather than one monolithic prompt. Each skill is a thin, agent-readable contract that wraps deterministic primitives, and the quality gates are hard pass or fail render checks instead of soft preference scores. Here is how a document moves through the pipeline:
- Paper2Assets reads the source once and extracts a shared, structured asset bundle: figures, tables, section text, and claims. Every downstream skill draws from this bundle, so nothing is re-parsed from scratch.
- Paper2Poster lays out a print-ready poster and runs a measured-fill loop that refuses to ship until load-bearing sections are actually populated, not left visually empty.
- Paper2Video produces a synchronized talk video with narration aligned to slides, generated from the same asset bundle.
- Paper2Blog writes a bilingual blog post that mirrors the poster and video so the three artifacts tell one consistent story.
- Paper2Reel is the interactive convergence layer, an HTML viewer that binds the poster, video, and blog into a single reviewable reel.
The pipeline runs on OpenAI Codex or Claude Code, and Microsoft lists Claude Opus 4.6 or higher and GPT-5.5 or higher as the supported model tiers. Because the skills are text contracts, they slot directly into the agent runners builders already use.
How It Compares to Manual Work and Prior Tools
The headline differentiator is editability. Earlier automation such as Paper2Poster and Paper2Video shipped one-way renders that an author could not reopen. ResearchStudio-Reel produces artifacts that round-trip through PowerPoint and Word, so a human can take the draft the last 10 percent of the way. The companion ResearchStudio-Idea paper details the same skill-contract philosophy applied to the front end of the workflow.
| Capability | Manual workflow | Single-artifact tools | ResearchStudio-Reel |
|---|---|---|---|
| Poster, video, and blog from one source | Hours per format | One format each, separate runs | All three in one pass |
| Shared extraction | Manual copy-paste | Each tool re-extracts | Paper2Assets, extracted once |
| Editable outputs | Yes, but built by hand | Mostly one-way renders | PowerPoint and Word round-trip |
| Consistency across artifacts | Depends on the human | Drift between tools | Enforced by shared assets |
| Quality control | Human review | Soft preference scores | Hard pass or fail render gates |
On aesthetics, the generated posters surpassed the authors' own original work across two held-out judges, and the overall win rate against prior systems and frontier models landed between 84 percent and 93 percent of papers tested.
Why It Matters for Creators and Builders
ResearchStudio-Reel is framed around research papers, but the pattern generalizes to any structured document a creator needs to broadcast: a product spec, a case study, a tutorial, or a report. What this enables is a repeatable content-production loop where one source of truth fans out into a deck, a video, and a written post that all agree with each other. That is the exact bottleneck small teams hit when they try to be everywhere at once with limited hours.
It is also a working reference architecture for agent skills. Anyone studying how to compose Claude agent skills into a production pipeline now has an open, MIT-licensed example with a shared extractor, editable generators, and measurable exit gates. Builders exploring similar orchestration should also see our coverage of the Claude Science AI Workbench, which tackles the research side with the same Anthropic tooling.
How to Try ResearchStudio-Reel
The repository offers two install paths. For a guided setup, run the interactive installer with npx github:microsoft/ResearchStudio, which walks through plugin selection and API key configuration. For a local clone, pull the repo and run install.sh to create the skill symlinks and set environment variables. You will need Python 3.10 or higher and access to a supported model through Claude Code or Codex.
Because the outputs are editable, the intended workflow is human-in-the-loop: let the pipeline generate the first full draft of your poster, video, and blog, then open them in PowerPoint or Word for the finishing touches. Anthropic's skills documentation is worth reading first if you want to understand the contract format the pipeline is built on. If you are already producing visual content from prompts, our look at Manim-Studio shows a complementary prompt-to-animation approach.
Frequently Asked Questions
Is ResearchStudio-Reel free and open source?
Yes. The code is published on GitHub under an MIT license, so it is free to use, modify, and build on commercially. You still pay for the underlying model calls through Claude Code or Codex.
Which models does it require?
Microsoft lists Claude Opus 4.6 or higher and GPT-5.5 or higher as the supported tiers, running through Claude Code or OpenAI Codex. It is designed to run on the same agent runners many builders already use.
Can I edit the poster and video after generation?
Yes, and that is the main advantage over earlier tools. The poster and slides round-trip through PowerPoint and Word, so a person can reopen and refine every artifact rather than accepting a locked one-way render.
Does it only work on academic papers?
The skills are tuned for papers, but the architecture operates on any structured document. A product spec, tutorial, or report can flow through the same shared extractor and generators, though results will vary outside the academic format it was benchmarked on.
How is it different from Paper2Poster or Paper2Video?
Those tools each produced a single artifact and re-extracted the source independently, which caused drift. ResearchStudio-Reel unifies extraction in Paper2Assets, delivers three consistent editable artifacts, and enforces hard render gates instead of soft preference scoring.
What does the 84 to 93 percent win rate mean?
Across the papers tested, ResearchStudio-Reel's outputs were preferred over prior systems and frontier language models on 84 percent to 93 percent of cases by held-out judges, and its posters beat the authors' own originals on aesthetics.
What to Do Next
If you regularly turn one document into multiple formats, clone the repository and run it on a piece you have already published to see how close the first draft lands. Study the skill contracts as a template for your own agent pipelines, then decide whether to adopt the full stack or borrow just the shared-extractor pattern for your existing tools.
