Super Research
Autonomous research skill for open-ended, high-volume research work, an agent left...
Install
npx skills add XiaomiMiMo/MiMo-Code --skill super-research -g -a claude-code -y- Repository
- XiaomiMiMo/MiMo-Code13.7k
- Category
- Testing
- Updated
- 3d ago
- License
- MIT
- Context cost
- Mediumabout 2,174 tokens
- Runs scripts
- Yes, review before use
What it does
Autonomous research skill for open-ended, high-volume research work, an agent left running for a while (minutes to overnight) that produces honest, comparable, auditable evidence instead of a single one-shot answer. Covers eight modes selected by the request: (1) experiment loop, iteratively edit code, run, measure a metric, keep or revert (baseline → hypothesize → run → keep/revert loop; use for "optimize X", "tune hyperparameters", "run experiments overnight", "autonomously improve this model", "hill-climb a metric", "自动实验"); (2) topic survey / 主题调研, collect and synthesize sources on a question (use for "survey the literature on X", "research topic Y", "调研 Z", "literature review", "deep research", "what's the state of the art in", "gather evidence about"); (3) quantitative analysis / 量化分析, reproducible, hypothesis-first data analysis with schema audit, effect sizes, and caveats (use for "analyze this dataset", "量化分析", "test whether X correlates with Y", "compute the effect of", "investigate this data"); (4) benchmark comparison / 对比评测, pick among N candidates under a fair, fixed matrix (use for "compare X vs Y", "which library/model/prompt is best for us", "benchmark these options", "选型", "对比评测"); (5) root-cause investigation / 根因排查, hypothesis-driven, two-way-reversal debugging of regressions, flakes, and perf drops (use for "why is X broken", "root cause this", "debug the regression", "why is it flaky", "排查", "定位", "复盘"); (6) ablation study / 消融实验, leave-one-out at
Autonomous Research
You are about to become an autonomous researcher. The value of this skill isn't the specific procedure, it's the property that research work done under it is comparable, honest, and auditable. A pile of ten cheap experiments/queries/analyses done to the same standard beats one clever untested claim. This is what makes it possible for a human to check on you eight hours later and actually trust what they see.
This skill was distilled from Karpathy's autoresearch methodology and generalized to six research modes.
Shared discipline (all modes)
Every mode operates under the same five rules. Read them before you branch into a mode.
-
State a contract before you begin. Infer everything you can from the workspace and the user's request, then write down what you're about to do, the goal, the primary output, the stopping condition, and get one confirmation. This is your last question. After confirmation you are autonomous.
-
Establish a baseline as your first artifact. In every mode there is a version of "the answer without any of my work", the unmodified code, the first three sources you find, the raw dataset before any transformation. Record it first. Without a baseline, "better" and "significant" are meaningless.
-
Every step is logged, in a machine-readable file, including failures. A tab-separated log (TSV, not CSV, descriptions contain commas) with a header row and one row per attempt. Failed attempts get logged with a
crash/dead-end/inconclusivestatus. Silently discarding attempts is the fastest way to fool yourself and the user; the log is the evidence that you actually did the work.
