research-workflow-runner

name	research-workflow-runner
description	Use when running this example repository's research workflow from a user research idea, producing active-node artifacts, handling step/options/choose/approve/reject, and continuing until research_complete.

Research Workflow Runner

This is a project-level skill for the example repository's research workflow. It is intentionally domain-specific. The reusable agent-harness-cli skill named harness-workflow-runner remains generic and should be used for arbitrary workflow-controller projects.

This repository has one workflow:

• Workflow spec: workflows/research.json
• Artifact contract: research/README.md
• Runtime context: research/context.json
• State path: .agent-harness/research-state.json
• Report directory: reports/research

Users should be able to start with a short prompt:

Research this idea with the research workflow:

<idea>

Read AGENTS.md first. It is the project source of truth for how to operate the workflow.

Operating Loop

1. If state exists, continue from the active node. If no state exists, start at research_context.
2. In research_context, create research/context.json and the matching research/<topic_slug>/ directory from the user's idea.
3. Read the active node's artifact contract in research/README.md.
4. Resolve {topic_slug} from research/context.json.
5. Create or update the required artifact under research/<topic_slug>/.
6. Include every required heading exactly as specified.
7. Include one routing line: Status: <allowed_status>.
8. Choose the status honestly based on evidence.
9. Expect every stage gate to run semantic Markdown checklist checks, and key research gates to run deeper quality checklists.
10. Maintain evidence-control artifacts when the active stage requires them: source_evidence_pack.csv, data_fitness_matrix.csv, claim_evidence_trace.csv, figure_blueprint.csv, figure_manifest.csv, strict_paper_review.md, reviewer_scorecard.csv, and quality_report.md.
11. Run or allow the Stop hook to run:

agent-harness step --task workflows/research.json --hook-json

If state is choosing, inspect options and choose:

agent-harness options --state .agent-harness/research-state.json
agent-harness choose <transition-id> --state .agent-harness/research-state.json --reason "<reason>"

If state is waiting, ask the user the required question, then use approve or reject with a concrete reason.

The quality gates are intentionally stricter than format checks. If a checklist blocks, repair the research logic rather than only editing prose.

Do not make unsupported claim-strength jumps. If the trace only supports a mechanism claim, do not write causal, market-share, price, or supplier-selection claims in the manuscript.

At figures_manuscript, generate real figure/table assets before setting Status: figures_synced. Save them under research/<topic_slug>/figures/, write figure_manifest.csv, and cite each asset ID/title/path in manuscript.md. Prefer Python for data-derived visuals and computed tables. Use imagegen only for conceptual mechanism visuals or graphical abstracts where exact numeric fidelity is not required.

Keep the main manuscript body publication-facing. In Abstract, Introduction, Methods, Results, Discussion, and Conclusion, do not mention workflow, checks, checklists, local paths, internal CSV artifact names, source IDs, claim IDs, or figure asset IDs. Put manifest paths only in Figures and Tables. Use reader-facing citations and include a References section.

At strict_paper_review, act as a strict external reviewer. Score novelty, literature positioning, evidence adequacy, method validity, analysis quality, claim calibration, figures/tables, writing, reproducibility, and references. Use Status: accept only when all score thresholds are met. If not, route back to the stage that can actually repair the issue.

Do not edit workflow specs or checks during normal operation.