Research arena
Skill hanxinzhang/research-arena/.agents/skills/research-arena
LLM-native research community protocol for agent-based research workflows
npx -y skills add hanxinzhang/research-arena --skill research-arenaAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use when the user wants to run or extend Research Arena, an LLM-native research community protocol with competing Researchers, an Integrity Checker, Referees, revision, and an Editor/Publisher decision. Trigger for agent-based research workflows, peer-review simulation, manuscript/revision rounds, OASIS demo runs, or adding new Researcher/Referee agents.
SKILL.md
2.6 KB, as published. Nobody here has run it
Research Arena
Use this skill to run or extend the Research Arena protocol from this repository. Research Arena is a repo-native workflow, not a Python package or command-line runner. The LLM agent is the runtime: it reads the protocol files, plays the named agent roles, creates artifacts, and coordinates review/revision.
Core workflow
- Read
program.md. - Read
AGENTS.md. - Read the participating agent folders under
agents/. - If the user asks for the OASIS demo, read
data/README.mdfirst. The raw CSV is not committed; stop and ask the user to create it locally if missing. - Use prompts in
prompts/when the user wants a ready-made start, continuation, or audit prompt. - Treat
examples/oasis_demo/as illustrative output, not as raw source data.
Operating rules
- Keep Research Arena LLM-agnostic; do not assume Codex is the only possible backend.
- Researchers compete using the same shared resources, but they do not need a shared target, focal variable, metric, or research question.
- Enforce the integrity gate before editorial acceptance.
- Do not treat outputs as clinical advice, diagnostic evidence, causal discovery, or validated science.
- Do not fetch, redistribute, or invent third-party datasets unless the user confirms the source and license/redistribution terms.
- Generated analysis code should avoid network calls, shell calls, credentials, private paths, hidden files, and destructive operations.
Common tasks
- Run a demo: Start from
prompts/start_oasis_demo.md. - Continue revisions: Start from
prompts/continue_revision.md. - Audit a run: Start from
prompts/audit_run.md. - Add agents: Copy from
agents/templates/researcher/oragents/templates/referee/, then updateconfig.json,profile.md, andrules.md. - Inspect examples: Use
examples/oasis_demo/summary.md,final_decision.md,integrity_report.md, andreferee_reviews.md.
Artifact locations
Follow the artifact contract in program.md:
runs/<run-id>/submissions/<run-id>/agents/<agent-id>/workspace/<run-id>/
Before publishing or sharing artifacts, audit for private data, raw data leakage, unsafe generated code, unsupported claims, and stale outputs.