文献筛选表

Use for requests such as “按中国企业实证研究筛选这50篇文献，给我Excel和排除理由”。 Reuse cnki-search/advanced-search, parse-results and paper-detail for website access; this skill does not operate a new crawler. Existing authorized scope carries through.

1. Freeze the observed result set and record query, date range, source platform, sorting, retrieval time and original rank. Define inclusion/exclusion criteria from the user's research question. If useful, propose a narrow criterion; do not silently restrict the population, date range or method.
2. Read the shared ledger schema. Use ../cnki-resume/scripts/workflow.py to init/import the task, or use the supplied existing task file. All three new skills require cnki-resume installed alongside.
3. Read each paper's actual title/abstract/full text as available. Store include, exclude or uncertain, a specific reason tied to criteria, a short exact evidence excerpt, the basis and access level using screen. Avoid numeric relevance scores that imply unsupported precision. Do not treat unavailable full text as irrelevant. A title-only exclusion needs an unambiguous criterion violation; otherwise mark uncertain. Uncertain records require follow-up.
4. Persist each decision immediately. Never replace the result set with more convenient papers. Retain exclusions and original ranks. Similar title/year matches flagged by import are review candidates, not automatic duplicates.
5. Export with export --output <new-folder>. Deliver 研究工作表.xlsx with the 文献筛选表 sheet and UTF-8 CSV fallback. Report counts of included/excluded/ uncertain/unprocessed records and the evidence-access limit.

Changing a decision invalidates that paper's extraction and records the old/new values in the event history. Screening criteria are frozen per task: create a new task for changed criteria, so earlier decisions are never silently reused.

Optional Jev proposals

If the user chooses paid Jev-assisted screening and cnki-jev is installed, read its SKILL.md. Installation alone never enables paid calls or abstract upload. The adapter reads the frozen task and returns per-condition proposals; it cannot finalize screening or supply source evidence. Shadow mode leaves baseline judgment independent. In assist mode verify the source evidence for inclusion proposals; route every exclusion, uncertain answer or conflict to the baseline LLM/human review. Save the final reason and exact quote via the same screen command, optionally with decision_provenance as described by cnki-jev. Missing/disabled extension uses the normal workflow with no Jev calls; report configured fallback.