Guide · Agent skill
How to use the super-research agent skill: setup, triggers and examples
Updated · written with AI from the project's own documentation and checked against our catalogue data
super-research is an agent skill for Claude that turns a long research task into a disciplined, logged process. Instead of one quick answer, the agent agrees on a plan with you, records a baseline, runs many small experiments, analyses or source checks, and writes every attempt to a file you can audit later. This guide explains what the skill does, how to install it, how to trigger it with plain requests, and what to watch for before letting it run unattended.
Super Research (Autonomous Research)XiaomiMiMo/MiMo-Code · MIT Download the skill Quick start
- 1Download the ZIP. Download the super-research ZIP from this page.
- 2Unzip into a skills folder. Unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ inside a project.
- 3Or upload in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the skill.
- 4Confirm the prerequisites. Make sure Python 3, bash and network access to arXiv, Semantic Scholar, OpenAlex and Crossref are available.
- 5Describe the research task. Ask in plain language, for example to compare three options or optimize a metric, and let the agent pick the mode.
- 6Confirm the contract. Read the goal, primary output and stopping condition the agent proposes, then confirm once.
- 7Review the log. Read the TSV audit log, including baseline and failed attempts, before trusting the result.
What the super-research skill makes the agent do
The skill is built around one property: research done under it is comparable, honest and auditable. The agent works through a fixed set of rules rather than a single one-shot answer. It states a contract (goal, primary output, stopping condition), gets one confirmation, then runs autonomously until the agreed stopping condition or until you interrupt it.
The skill covers eight modes, and the agent picks one from your request: an experiment loop that edits code, runs it, measures a metric and keeps or reverts the change; a topic survey that collects and synthesizes sources; quantitative analysis of a dataset; benchmark comparison among several candidates; root-cause investigation of regressions and flaky tests; ablation studies of system components; paper reproduction; and paper writing with citation auditing.
- Experiment loop: baseline, hypothesize, run, keep or revert
- Topic survey: collect and synthesize sources on a question
- Quantitative analysis: hypothesis-first analysis with schema audit, effect sizes and caveats
- Benchmark comparison: pick among N candidates under a fixed matrix
- Root-cause investigation: hypothesis-driven debugging of regressions, flakes and performance drops
- Ablation study: leave-one-out attribution against a measured noise floor
- Paper reproduction: implement a paper's method as a working repo, logging ambiguities
- Paper writing and citation audit: verify citations against real API records
Install the super-research skill
You can install the skill in two ways. For a personal setup, download the ZIP from this page and unzip it into ~/.claude/skills/. To scope it to one project, unzip it into the .claude/skills/ folder inside that project.
If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.
- Personal install: unzip the downloaded ZIP into ~/.claude/skills/
- Project install: unzip the downloaded ZIP into .claude/skills/ in your project
- Claude apps: Settings → Capabilities → Skills, then upload the skill
Check what you need before running it
The skill needs Python 3, because the bundled benchmark scripts call python3, and bash, because the toy benchmark runner run_case.sh is a bash script. It also needs network access to free scholarly APIs: arXiv, Semantic Scholar, OpenAlex and Crossref. According to SKILL.md, no API keys are required.
The skill runs scripts and uses the network, so treat it as code execution on your machine. Run it in a sandbox or a copy of your repository when the work touches anything you care about.
- Python 3 installed (the benchmark scripts call python3)
- bash available (run_case.sh)
- Network access to arXiv, Semantic Scholar, OpenAlex and Crossref
- No API keys required according to SKILL.md
Trigger the skill with natural requests
You do not need a special command. Describe the work in plain language and the agent selects the matching mode. The skill also triggers on research work with volume and discipline, even if you never use the words 'research' or 'experiment'. Below are example phrasings for each mode, taken from the skill's trigger list.
If your request is ambiguous, the agent picks the closest mode and states its choice in the contract. Check that statement before you confirm.
- Experiment loop: "optimize X", "tune hyperparameters", "run experiments overnight", "hill-climb a metric"
- Topic survey: "survey the literature on X", "literature review", "what's the state of the art in"
- Quantitative analysis: "analyze this dataset", "test whether X correlates with Y"
- Benchmark comparison: "compare X vs Y", "which library is best for us", "benchmark these options"
- Root cause: "why is X broken", "debug the regression", "why is it flaky"
- Ablation: "ablate X", "which parts of Y matter", "is component Z pulling its weight"
- Paper reproduction: "implement this method", "reproduce the main table of X"
- Paper writing: "write a paper on X", "polish this draft", "citation check", "detect fabricated references"
Example: a benchmark comparison and what gets logged
Suppose you ask the agent to compare three text-cleaning implementations and tell you which one to use. The skill selects benchmark comparison mode. The bundled toy benchmark shows the expected behavior: run each of the 3 candidates on each of the 5 fixed cases with at least 2 repetitions, log every cell including failures, and produce a ranking with a drop-one-case stability check.
The toy benchmark's runner takes a candidate and a case and prints a similarity score from 0.0 to 1.0, where higher is better, plus a wall-clock time. If a candidate crashes or times out, the harness prints status=error and score 0. The log is a tab-separated file, so every attempt, including failures, stays visible.
- Each attempt is one row in a TSV log with a header row
- Failures are logged with a crash, dead-end or inconclusive status
- The comparison holds the test set fixed rather than tuning toward a favorite
./run_case.sh <candidate> <case>Read the audit trail and review the results
The main value of the skill is that you can check the work later. The agent does not stop to ask permission mid-run, so the log is what you read the next morning. Look for the baseline row first, then the attempts, then the failures. A clean log that includes dead ends is a good sign: the agent did not silently discard attempts.
Treat the conclusions as the agent's proposal. The skill itself says the human decides later, and citation checks only cover what the free APIs return.
- Start with the baseline row to see what 'without any of my work' meant
- Check that failed attempts are present, not removed
- Verify any citation or ranking against the raw evidence before you rely on it
Safety notes before an unattended run
The loop keeps going without checking in. It can edit code or run commands for hours, so do not point it at production systems or shared repositories without a sandbox and a clear stopping condition. Agree on the stopping condition in the contract step, because that is the only point where you are asked anything.
The skill also says never to game the metric, sources or analysis. It should not edit eval code because it looks wrong, drop sources that do not fit a thesis, or cherry-pick a nicer subset. If it suspects a problem with ground truth, it logs the concern and continues.
- Use a sandbox or a copy of the repo for experiment loops
- Set a clear stopping condition before confirming the contract
- Not suitable for final academic claims without human review
What people use it for
- Run an overnight experiment loop that tries code changes against a measurable metric and keeps only improvements
- Survey the literature on a question and collect the sources with a traceable record
- Compare three libraries or models on a fixed set of test cases with a ranking
- Debug a flaky test or performance drop by testing hypotheses and logging each one
- Check the citations in a draft paper against records from free scholarly APIs
Frequently asked questions
Is the super-research skill free?
The skill's source is published under the MIT licence. It uses free scholarly APIs (arXiv, Semantic Scholar, OpenAlex, Crossref), and SKILL.md states that no API keys are required.
How do I install super-research for Claude?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ inside a project. In the Claude apps, you can upload it under Settings → Capabilities → Skills.
What does super-research need to run?
It needs Python 3, because the bundled benchmark scripts call python3, and bash for run_case.sh. It also needs network access to arXiv, Semantic Scholar, OpenAlex and Crossref.
Does super-research need API keys?
According to SKILL.md, no API keys are required. It uses the free scholarly APIs listed in the skill.
Is super-research safe to run on my codebase?
It runs scripts and uses the network, and the experiment loop can edit code and run commands for hours without checking in. Use a sandbox or a copy of the repository and set a clear stopping condition first.
How is super-research different from a normal chat answer?
Instead of one answer, it works through a contract, a baseline, and many logged attempts, including failures, in a TSV file you can audit. It is designed for research work with volume and discipline, not quick factual questions.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use super-research if you want an agent to run a long, logged research task such as an overnight metric search, a literature survey, a fair benchmark, or a citation check, and you can review the audit trail afterwards. Skip it for quick factual questions, unattended work on production systems or shared repos, and final academic claims that have not been reviewed by a human.
Download the skill