Refine Literature Question
Discipline And Interview Setup
Read discipline-routing.md, then only the selected discipline profile. Reuse the user's existing brief; do not restart intake at each stage. For journalism/communication or information science, use that profile's concepts, extraction fields, and appraisal checks instead of a generic social-science template.
When the user requests "grill me", 追问, or 模拟答辩, read research-interview.md. Ask one decision-relevant question per turn and wait; do not generate the stage deliverable until enough decisions are settled or the user asks to proceed. In quick/no-question mode, record assumptions and continue without inventing answers.
For evidence extraction or drafting, read review-workflow.md. With no sources, give a plan or outline, not a purported synthesis. If reference files are unavailable, retain the source guardrails below and state that discipline guidance is missing.
Goal
Turn a broad topic into a usable question anchor before searching, reading, or writing the review. The output should be a one-sentence research question that forces selection: some papers become clearly relevant, and others become safe to skip.
Workflow
- Identify whether the user's input is a topic, a phenomenon, or an actual question.
- Identify the review type, then replace vague nouns with bounded concepts, objects, relationships, periods, or puzzles.
- Generate 3-5 candidate angles after scope decisions are clear; do not flood an interview turn with candidates.
- Evaluate each candidate against the three-anchor test.
- Recommend one primary question and 1-2 backups, then explain what each would make the review include or exclude.
Three-Anchor Test
A strong literature review question should:
- Match the review purpose: explanation, bounded evidence mapping, conceptual distinction, or historical interpretation. A causal "why/how" question is not mandatory.
- Force literature selection by making some studies irrelevant to the question.
- Feel worth answering if a reader could get a credible answer from the review.
If the question fails any item, revise it before proceeding to search or outline the literature.
Question Patterns
Use these patterns when generating candidates:
- Mechanism:
Why does [phenomenon] occur in [context], and through what mechanism? - Contrast:
Why does [outcome] differ between [group/place/time A] and [group/place/time B]? - Condition:
Under what conditions does [factor] influence [outcome]? - Tension:
Why do existing studies disagree about [relationship/result]? - Process:
How does [actor/system] move from [state A] to [state B]? - Scoping:
What concepts, methods, and populations have been studied for [object] within [scope]? - Conceptual/history:
How has the meaning of [concept] changed across [traditions/periods], and what does that change explain?
Output Format
Provide:
markdown## Candidate Questions 1. ... 2. ... ## Recommended Anchor [One sentence] ## Why This Works - Selection boundary: ... - Likely literature modules: ... - What to skip: ... ## Revision Notes - If the user wants broader scope: ... - If the user wants narrower scope: ... ## 待核实清单 [Unverified sources, "未查到" results, sentences marked [需补来源], and inferences that still need checking; write "无" if empty] ## 给作者的问题 [Judgment calls that belong to the author; write "无" if empty]
The output sections above, and the two closing sections, keep fixed names and order; see references/output-contract.md (a copy bundled in this skill folder).
Guardrails
- Do not accept an unbounded label such as
XX research reviewas the final anchor. Descriptive, historical, and conceptual reviews are valid when their corpus, scope, and analytical purpose are explicit. - Do not invent a domain-specific claim when the user has not provided enough context; offer candidate angles and mark assumptions.
- Treat AI as a sparring partner: generate options and pressure-test wording, but keep the user's academic judgment in charge.
- Source integrity: follow
references/source-integrity.md. Claims about what "existing studies" do or disagree on are assumptions until the user supplies sources; mark them. Use only sources the user provided or that are verified, tag each with a status (用户提供 / 已核验 / 未核验), write "未查到" instead of completing authors, year, journal, or pages, and put anything unverified in## 待核实清单. If that file is not available, still apply these rules.
