research-sop

Research SOP

Use this skill as the main router for empirical research. Custom researcher, engineer, verifier, reviewer, and budgeter agents are optional helpers; when they are unavailable, execute the same role boundaries inline.

Inputs and promised output

Before doing expensive work, establish:

• the research question and decision the result should support;
• the active workspace and relevant files or datasets;
• whether the work is exploratory or confirmatory;
• cost, time, reserved-data, and external-service constraints;
• whether the user asked only for diagnosis or also authorized implementation.

Finish with a research handoff containing the question, literature basis, hypothesis node ids, comparison evidence, experiment manifest, verification status, unresolved risks, and the next justified action. Do not call a result publication-ready merely because code ran or unit tests passed.

1. Orient and resume

1. If Cockpit is useful, tell the user how to start it from the same workspace: claudescientist cockpit --workspace <path> --lang zh.
2. Inspect the active frontier and ancestors. Reuse an existing question or hypothesis when it already represents the task; otherwise create the question and candidate hypothesis nodes.
3. Call mcp__memory__match_signatures with the task and likely failure modes. Read matching failures before choosing a method.
4. Query existing literature and contradictions. Record a snapshot before a risky pruning decision or when resuming an interrupted investigation.

Expected state: a question node or explicit task statement, known prior failures, and a short list of missing evidence.

2. Establish the evidence base

1. Call mcp__memory__query_literature, find_baselines_for, and find_contradictions as relevant.
2. Use arXiv or OpenAlex only when those optional MCPs are available. If they are absent, continue with repository sources and say what literature search remains incomplete; do not invent citations.
3. Ingest structured paper summaries that materially affect the decision.
4. Separate established facts, reported claims, assumptions, and open gaps.