Research Companion: a plugin that argues with you about which paper to write

The premise is that problem selection is where research goes wrong

Most AI writing tools help you write papers, and this one aims at the earlier question, which paper to write. The README frames the gap in terms of a missing colleague. Researchers, it says, do not lack the ability to write papers, they lack someone trusted who will tell them an idea is not worth six months of their life before they invest those months, ask who else is working on the same thing and what their unfair advantage is, challenge them to state the key insight in one sentence, help them find the unexpected cross-field connection, and evaluate whether a struggling project should be continued, pivoted or killed. The inspiration is Nicholas Carlini's essay on winning a best paper award, which the README credits with the argument that great research starts with taste, strategic problem selection, honest self-evaluation and knowing when to kill your darlings.

Three agents with different jobs, and one shared skill

The agent set is small and the roles are distinct. The Idea Critic stress-tests research ideas along seven dimensions: novelty, impact, timing, feasibility, the competitive picture, the nugget, and narrative potential, and it returns a Pursue, Refine or Kill verdict. The Research Strategist works at project level rather than idea level, doing triage of continue, pivot or kill, comparative advantage mapping, impact forecasting, opportunity cost analysis and scooping risk assessment. The Brainstormer is described as an enhanced creative brainstormer with an explicit focus on cross-field connections, on what the README calls strategic ignorance, which means challenging flawed assumptions the field follows uncritically, and on the skeptical-reader test. Above them sits a single skill, invoked as /research-companion, that orchestrates all three agents through a structured multi-phase session and includes Carlini's conclusion-first test. There is also a set of 8 research strategy principles in three categories, Problem Selection, Execution Strategy and Strategic Positioning, that guide how the agents evaluate.

The six-phase session and the conclusion-first test

The skill runs in a fixed order, and the order is the interesting part. Seed understands your problem space, your interests and what bugs you about the field. Diverge generates ideas, alternative framings and cross-field connections. Evaluate stress-tests the top two or three ideas with the Idea Critic. Deepen checks novelty, positioning and the competitive picture. Frame writes the abstract and conclusion as if the paper were already done. Decide gives a final assessment with next steps, starting with the single riskiest assumption to test first. The conclusion-first test lives in the Frame phase and is stated bluntly in the README: if you cannot write a compelling conclusion now, the idea is not ready. That is a cheap filter with real value, because a paper whose conclusion you cannot state is a paper whose contribution you have not yet found.

Installing it as a Claude Code plugin

The installation is two commands, and the plugin comes from a marketplace.

claude plugin marketplace add https://github.com/andrehuang/research-companion
claude plugin install research-companion@andrehuang-research-companion

The first command registers the repository as a marketplace source and the second installs the plugin from it. The repository layout supports this. There is a .claude-plugin directory, an agents directory holding the agent definitions, a skills directory, a principles directory for the strategy principles, a codex-skills directory, a docs directory, an assets directory, a marketplace.json file, and a .agents directory. The skill is triggered with a slash command inside Claude Code, in the form /research-companion followed by your question. The README also shows the plugin working in a more conversational mode, where you describe an idea in plain text and ask for an evaluation, and the Idea Critic answers with a verdict and the single most important question to resolve next.

Codex support, and a repo-scoped marketplace to avoid manual JSON

The repository now includes native Codex support, and the details matter if you work in that ecosystem. The manifest files are .codex-plugin/plugin.json, agents/openai.yaml and .agents/plugins/marketplace.json. The README states that the same core prompts, agents and principles are shared across Claude Code and Codex, and that the orchestration instructions are written to degrade gracefully if delegated subagents are unavailable, which is a sensible way to handle a tool that assumes a multi-agent runtime. Codex supports plugin marketplaces and installation through a Plugin Directory, and the repository adds a repo-scoped marketplace file so the plugin can be installed after cloning without manually creating marketplace JSON. Setup details for Codex live in a separate installation document in the docs directory. In Codex the skill is invoked with a dollar sign prefix rather than a slash.

Prompt definitions, not an application

It is worth being blunt about what this project is, because the README is not a program description. The primary language is listed as unknown, and the repository is a collection of prompt and metadata files. There is no model to serve, no database, no retrieval index and no API. The evaluation in the Idea Critic is reasoning performed by whatever agent platform you install the plugin into, following the instructions in these files. That has direct consequences for what the seven-dimension verdict is worth. It is a structured, repeatable way of being talked out of a bad idea, not a novelty check against published work, and the dimension list makes that clear. Novelty is asked as a question you must answer, how long until someone else does this, rather than something the tool determines by searching a corpus.

Now folded into a larger package, with maintenance spread across two repos

The April 2026 update in the README is the most consequential fact for anyone planning to depend on this. Research Companion is now also a component of researcher-pack, described as a more advanced, all-in-one Claude Code plugin that bundles strategic brainstorming, paper reading, literature tracking, academic writing and project management into a single integrated workflow. The README claims the standalone repository is still maintained, with updates synced between both repositories, though the last push dates to April 2026, so you can use it here on its own or take the full toolkit from the other project. In practice, a new user starting today is probably better served by the all-in-one package, and the standalone repo is the narrower option for people who only want the strategic layer. The licence is MIT, and the repository is not archived, but the last push was on 2026-04-13 and there are no GitHub releases, so there is no version history to pin against.

Editorial conclusion

Research Companion fits a researcher who has more ideas than time and wants an adversarial check before committing six months to one, and who is willing to take the Kill verdict. It does not fit anyone expecting literature coverage, novelty checks against a database, or citation data, since the evaluation is reasoning over prompts rather than retrieval. Verify first that your agent platform can load the plugin, read the dimension table before you treat a verdict as objective, and check the repository's release state, since the last push was on 2026-04-13 with no tagged releases and a separate all-in-one package now carries most of the functionality.

Frequently asked questions

Official sources

1. andrehuang/research-companion on GitHub
2. Issues
3. License: MIT
4. README