0-autoresearch-skill_skill

Overview

This skill orchestrates end-to-end autonomous AI research projects using a two-loop architecture. It manages workspace state, runs fast inner-loop experiments and periodic outer-loop synthesis, and routes execution to domain-specific skills. It also supports continuous operation via /loop and heartbeat mechanisms and produces progress reports, presentations, and papers.

How this skill works

The skill bootstraps a project with a structured workspace and research-state.yaml, then executes a repeating two-loop cycle: rapid inner-loop experiment iterations with locked protocols and measurable targets, and outer-loop reflections that synthesize results and steer direction. For execution it routes tasks to specialized domain skills (data, training, evaluation, interpretability, infrastructure). It maintains logs, findings, and experiment artifacts and can run continuously using Claude Code /loop and OpenClaw heartbeat.

When to use it

• Starting a new research project from idea to paper
• Running many rapid experiments that need automated orchestration
• Managing multi-hypothesis or multi-condition studies
• Maintaining continuity across long-running experiments with heartbeats
• Producing progress presentations or draft papers from accumulated findings

Best practices

• Initialize the canonical workspace and research-state.yaml before running experiments
• Form testable, metric-driven hypotheses and lock protocols (commit before run)
• Keep inner-loop experiments short and measurable; reflect after ~5–10 runs or on notable patterns
• Use outer-loop reflections to synthesize findings.md into a coherent narrative before broadening or pivoting
• Route execution to appropriate domain skills instead of reimplementing low-level steps
• Record sanity checks, baselines, and temporal experiment metadata for reproducibility

Example use cases

• Optimize model hyperparameters with repeated short training runs and track an optimization trajectory
• Test mechanistic hypotheses (causal interventions, interpretability probes) with confirmatory protocols
• Run multi-hypothesis ablation studies and synthesize negative and positive results into a publishable narrative
• Maintain continuous autonomous experiments across days using /loop ticks and HEARTBEAT.md for agent continuity
• Generate interim progress presentations and a final paper draft from findings.md and experiment artifacts