lamm-mit/skills/research-pipeline

Research Pipeline Skill Overview

This is an orchestrator skill that manages a complete ML research workflow without performing the actual research tasks itself.

Core Identity

The orchestrator is a scheduler and validator, not a researcher. It:

• Checks for output files
• Reads summaries from prior phases
• Dispatches work to sub-agents via sessions_spawn
• Validates deliverables

As stated: "你不分析论文...你不写代码" (does not analyze papers, does not write code).

Key Execution Rule

Sequential, single-dispatch constraint: Each response can call sessions_spawn at most once. No parallel task launching. The orchestrator must wait for sub-agent completion before advancing to the next phase.

Six-Phase Pipeline

1. Literature Survey → papers/_meta/ directory with JSON files
2. Deep Survey → survey_res.md with method comparisons
3. Implementation Plan → plan_res.md with 4 sections (Dataset/Model/Training/Testing)
4. Implementation → project/run.py and ml_res.md with results
5. Review → judge_v*.md with PASS/BLOCKED verdict (up to 3 iterations)
6. Full Experiment → experiment_res.md with ablation studies

Task Dispatch Format

sessions_spawn requires:

• task: Starts with /skill-name, includes workspace path, context summary (2–5 lines), and expected output
• label: Phase identifier
• runTimeoutSeconds: Recommended 1800

The tool is called after reading prior outputs to bridge context between independent sub-agent sessions.