research-pipeline-runner_skill

Overview

This skill runs end-to-end research pipelines that create and manage workspaces, units, and checkpoints to produce survey, tutorial, snapshot, systematic review, or peer-review outputs. It scaffolds a reproducible workspace under workspaces/<name>/ with the required control files and artifacts. The runner enforces checkpoints so human approval is required before long-form prose is produced.

How this skill works

Given a one-sentence user goal (optionally with an explicit pipeline or constraints), the runner selects a pipeline and initializes a workspace with GOAL.md, PIPELINE.lock.md, UNITS.csv, CHECKPOINTS.md, and supporting files. It executes units one at a time, validates declared outputs against acceptance criteria, and writes status into UNITS.csv. The run halts for any HUMAN checkpoint and records approvals or decisions in DECISIONS.md before continuing.

When to use it

• You want a single command to create and run a reproducible research workspace end-to-end.
• You need evidence-first outputs (units, artifacts, acceptance checks) rather than immediate long prose.
• You want to run an arXiv survey, LaTeX survey, tutorial, snapshot, systematic review, or peer review pipeline.
• You need the run to stop for human approvals at scope/outline or other checkpoints.
• You need offline import support or controlled network use for arXiv/PDF steps.

Best practices

• Provide a clear one-sentence goal and optional pipeline name to avoid ambiguous pipeline selection.
• Never request automatic prose generation past HUMAN checkpoints; approve explicitly in DECISIONS.md when ready.
• Keep UNITS.csv as the single execution contract and only mark DONE when outputs meet acceptance criteria.
• Use strict mode when you want scaffold artifacts treated as stubs until manually refined and marked refined.ok files.
• Run network-heavy steps only when you allow external access; use offline imports where network is restricted.

Example use cases

• Kickoff an arXiv Markdown survey workspace and stop at C2 for outline approval.
• Run an arXiv LaTeX pipeline that builds papers/outline/latex and pauses for human sign-off before full prose.
• Execute a tutorial pipeline that sequentially generates sections, with unit-level acceptance checks.
• Resume a paused workspace after adding DECISIONS.md approval and continue automated units.
• Use strict mode to block scaffold JSONL from being promoted to final prose until manual refinement markers are added.