gaasher/Agent-Loop-Skills

Alpha Evolve evolves a model or program with parallel SEARCH/REPLACE mutations, cascade-evaluated runs, and a MAP-Elites archive across islands. Use it for bounded, diversity-preserving search over code or ML experiments rather than a single refine loop.

Anomaly Investigation diagnoses a known data anomaly by forming candidate causes, testing them against the data, and eliminating those the evidence refutes. Use it when you already have a spike, drop, or outlier and need the confirmed root cause and supporting evidence.

Blue Team patches a target against a concrete set of failing cases, one root-cause class at a time, while checking that previously passing cases still pass. Use it for red-team failure catalogues or CI test failures when you want the fix loop to stop only when regressions are closed.

Claim Verify Loop checks each discrete claim in a results draft against the underlying dataset, then stress-tests it for outliers, confounds, and subgroup effects. Use it before publishing when a data-backed draft needs adversarial verification and revision.

Data Analysis Loop performs iterative exploratory analysis on a dataset, testing one hypothesis at a time and only keeping findings that reproduce with a meaningful effect size. Use it for open-ended discovery when every claim needs a computed number behind it.

Dueling Autoresearch runs two different approaches against the same metric in parallel and keeps a shared scoreboard. Use it when you want an analysis-first head-to-head between lanes such as classical versus learned methods.

Exploratory Autoresearch runs an autonomous ML research loop that alternates analysis with scheduled swings, merges, and exploits. It is used for open-ended experiments where you want broad exploration first and a stagnation guard to prevent getting stuck in small-step tuning.

Hypothesis Generation Loop generates candidate research hypotheses and vets them against retrieved literature. Use it when you need a ranked pool of novel, testable hypotheses rather than a single research question or a proposal review.

Karpathy Autoresearch runs a fully autonomous loop that edits training code, executes the run, and keeps changes only when a single scalar metric improves. Use it for hands-off ML experimentation where the agent loops until interrupted.

Literature Search provides a shared CLI for discovering papers, checking novelty, searching full text, and walking citation graphs. Use it when a loop needs scholarly literature or experimental-result extraction and should fall back cleanly to web search if needed.

Literature Survey Loop builds an evidence-and-contradiction matrix for a research question by iteratively searching, extracting claims, and mapping source stances. Use it for structured literature surveys that need saturation, dispute tracking, and citation-backed coverage.

ML Autoresearch Loop runs an autonomous machine learning research cycle where each experiment is analyzed before the next change is made. Use it for open-ended model improvement with one change per run, optionally grounded in literature.

Optimize Loop runs an evaluator-optimizer cycle that keeps a correctness gate passing while reducing a measured cost. Use it to refactor code or tune SQL when results must stay unchanged but complexity or latency should drop.

Plan Loop turns a coding or engineering prompt into an executable plan with PR-sized tasks, dependencies, tests, and subtasks. Use it before implementation when a downstream agent or smaller model needs a clear handoff.

Power Analysis Loop sizes a single two-arm comparison by simulation, checks the design for validity flaws, and writes a preregistration before data collection. Use it for A/B tests, simple RCTs, or two-model evaluations with one primary outcome.

Prompt Optimization Loop improves an existing prompt by making one targeted edit at a time and re-running a user-provided eval command to see whether the score rises. It is for prompts that already have a measurable metric, not for writing prompts from scratch or changing model settings.

Purple Team runs red-team and blue-team as separate phases to harden a guardrail, classifier, prompt, or API you own. It repeats find, fix, and re-verify cycles until a fresh attack pass is dry or the cycle budget is reached, then opens a pull request.

Red Team stress-tests a guardrail, classifier, prompt, or API you own or are authorized to test. It generates adversarial inputs, compares target results against an oracle, and catalogues distinct failure classes with reproducible examples.

Research Proposal Loop iteratively grades and revises an existing research proposal using literature-grounded feedback. Use it when a proposal needs stronger soundness, contribution, and evidence before it passes a target score.

Research Question Loop sharpens a vague topic into a few strong research questions. It drafts candidates, scores them with a Specific/Answerable/Novel/Feasible/Significant rubric, and revises weak areas until enough clear the bar.

SWE Loop turns a coding prompt into implemented, tested changes by running a plan loop, then iterating Engineer and QA subagents until each task passes tests and quality checks. Use it for features, fixes, or refactors that need regression-safe delivery in a real repository.

Scientific Figure Loop generates a publication-style scientific figure from data or a brief and iterates on it with an adversarial critic. Use it when you need a figure rendered from scratch, checked against a fixed rubric, and revised until it meets a quality threshold.

Scientific Writer Loop revises an existing scientific draft with its data, figures, and optional code until it meets a quality threshold. It uses specialist judges plus an independent peer reviewer to critique, grade, and regenerate prose, figures, and code.

Tabular Cleanup rewrites messy CSV, TSV, Excel, JSON, or parquet data to match an inferred contract of binary checks. Use it when you need iterative, auditable cleanup with a replayable pipeline and deterministic stop conditions.

Tournament Autoresearch runs a multi-agent ML research loop where several ideas compete, a judge ranks them against a rubric, and the winner is the only change tried. It is used for open-ended experimentation when you want to pressure-test hypotheses before spending compute.