CCF Literature Searcher

Invocation Controls

CCFA Handoff Mode: PARTIAL (Recommended). Follow metadata.ccf_skill_controls.handoff_question_mode and ../ccf-common/references/handoff-modes.md. Use ../ccf-common/references/routing.md to keep literature search separate from idea optimization, manuscript writing, experiment design, paper review, and rebuttal.

Load ../ccf-common/references/task-modes.md before deciding exploratory, quick, or standard mode. Use exploratory mode for early direction scouting, "看看还有没有机会", "这个方向是不是被做完了", or literature search meant to feed idea optimization rather than a final novelty verdict. Use quick mode for a narrow related-work scan or a small set of candidate citations. Use standard mode for Related Work, Introduction, mature idea novelty grounding, benchmark discovery, experiment design, or substantial multi-deliverable work. A small citation lookup stays quick even if a writer will use it.

If the user asks for recurring watch, latest-paper monitoring, competitor tracking, "recently any similar idea", arXiv/OpenReview feed scans, or lab/project tracking, route to ccf-literature-monitor by the shared handoff mode. Use this skill for deep retrieval, closest-work clustering, related-work structure, benchmark/dataset discovery, and citation candidates.

Treat user ideas, draft text, unpublished results, and private manuscripts as private material. Load ../ccf-common/references/privacy-and-evidence.md before browsing. Search with public keywords, public titles, venue names, method names, public abstracts, or user-approved query text. Do not paste private draft sentences into a search query unless the user explicitly authorizes it.

Source-quality exclusion: do not search, cite, recommend, or include policy-excluded venues, journals, URLs, or PDFs. The shared policy includes MDPI sources in this exclusion set; record exclusions only in internal screening or the search-notes file.

Core Rule

Ground novelty and positioning using high-quality, inspectable sources. Prefer influential conferences, strong journals, official proceedings pages, archival repositories, and public paper pages. Do not invent papers, citations, venues, links, acceptance status, benchmark status, or numerical results. Separate searched evidence from inference. Literature search is not a kill gate: the presence of related work should produce differentiation options, open gaps, benchmark/evidence choices, and caution labels before any "direction is covered" conclusion. Follow the user's requested output shape: short list, related-work clusters, opportunity map, BibTeX candidates, benchmark table, search folder, or handoff summary.

Mandatory Search Checklist

In standard mode, complete this checklist before final output. In quick mode, run the relevant subset and return a compact checklist status.

1. The user's topic is converted into safe public search queries.
2. Shared source-quality exclusions are applied to search domains, candidates, and final outputs.
3. Sources prioritize primary or high-confidence venues: official proceedings, arXiv/OpenReview when appropriate, ACL Anthology, CVF, PMLR, ACM, IEEE, USENIX, DBLP, Semantic Scholar, OpenAlex, Crossref, and venue or project pages.
4. Candidate papers are deduplicated by title and linked to a stable URL.
5. Each included paper has venue/year/source status, paper type, and relevance rationale.
6. Score paper quality on insight, completeness, and experimental numeric evidence only when requested or useful for substantial screening, and only to the extent the inspected text supports it. Pure benchmark papers skip the numeric-results score and receive a benchmark-quality note instead.
7. Paper type is one of pure benchmark, pure method, method + benchmark, survey, system/tool, theory/proof, or other.
8. Every claim about a paper is traceable to the linked source or marked as inferred.
9. For idea-stage searches, each closest-work cluster includes what is already covered, what remains under-tested, and at least one possible differentiation or rescue route.
10. A literature-search folder is written when file access is available and the user asked for a reusable report or standard workflow.
11. When the search feeds idea optimization, an idea-grounding packet separates source-supported observations from inferred gaps and includes mechanism primitives, protocol anchors, cross-source conflict/open-gap relations, and confidence.
12. Optional handoff to ccf-literature-monitor, ccf-paper-writer, ccf-idea-optimizer, ccf-idea-reviewer, ccf-experiment-designer, or ccf-paper-reviewer follows CCFA handoff mode.

Workflow

1. Identify the search purpose: Related Work, Introduction support, novelty check, direction scouting, idea optimization, idea review, experiment design, benchmark/dataset discovery, or reviewer-risk diagnosis.
2. Create public queries from the user's topic. If the topic is too private or underspecified, ask only for non-sensitive keywords or infer broad keywords with lower confidence.
3. Load references/search-and-scoring.md. Use these search ranges as starting budgets, not quotas. Stop when the requested clusters, closest competitors, and evidence needs are covered and further retrieval adds no material information; broaden when a specific gap remains. Search breadth defaults:
   • Exploratory: 10-20 screened candidates, 5-10 final papers or clusters, plus opportunity gaps.
   • Quick: 6-10 screened candidates, 3-6 final papers.
   • Standard: 15-30 screened candidates, 8-15 final papers unless the user requests another size.
4. Search discovery indexes first, then verify candidates through stable paper pages or official proceedings when possible. Use broad web search only to find primary links; do not rely on snippets for final claims.
5. Filter by influence and fit. Prefer CCF-A/B conferences, top-field conferences, strong journals, widely used benchmarks, or recent high-signal preprints from credible groups. Exclude low-quality, predatory, inaccessible, or policy-excluded sources. For exploratory searches, include one or two "near miss" or negative-signal clusters if they reveal an open gap, failed assumption, outdated benchmark, missing user group, or neglected system constraint.
6. Classify papers as needed for screening. Score inspected papers only when requested or decision-relevant, using these dimensions:
   • insight: how clear and non-obvious the central idea is.
   • completeness: method/evaluation/proof/dataset/reproducibility coverage.
   • experimental numeric evidence: strength and relevance of reported numerical evidence; mark N/A benchmark for pure benchmark papers.
7. Write files only when reusable output is requested or needed within the authorized workflow. Honor no-new-files constraints. Reuse an existing canonical search folder for an update; otherwise use references/report-template.md with the default folder name:

output/literature-search/<topic-slug>/
  papers.md
  papers.csv        # only when structured reuse/export is needed
  search-notes.md    # only when queries/coverage must persist separately
  idea-grounding.md  # include when the search feeds idea optimization

Reuse an existing dated or custom search folder instead of renaming it. Store the current search date and source versions inside the report; do not create a new folder for an ordinary update. Derive requested Markdown/CSV views from one screened source set rather than writing independent copies. Skip empty or unrequested score columns and generic status sections.

8. If the search feeds another module, provide a compact handoff with canonical source paths and only decision-relevant evidence:
   • For writing: closest-work groups, novelty gaps, citation cautions.
   • For idea optimization: a compact idea-grounding packet with evidence cards, mechanism primitives, protocol anchors, gap/conflict relations, stale/overcrowded directions, timely pivots, and minimum viable research questions. Keep whole abstracts and generic background out of the handoff.
   • For idea review: novelty confidence and likely prior-art risks.
   • For literature monitoring: watch queries, tracked competitors, and recurring overlap signals.
   • For experiment design: datasets, baselines, metrics, benchmark protocols.
   • For paper review: missing related work and baseline risks.

Adaptive Output Contracts

Return the requested artifact first. If the user asks for a list of papers, output the list/table directly. If they ask for Related Work material, output clusters and positioning notes. If they ask for a folder, write the folder and summarize it. Use the following defaults for standard or quick search reports.

For standard search, return:

Search purpose:
Queries used:
Source policy:
Folder written:
Top paper table:
Closest-work clusters:
Opportunity map:
Quality-score rationale:
Benchmark/dataset candidates:
Novelty and positioning risks:
Recommended next module:
Checklist status:

For quick search, return:

Quick search scope:
Top candidates:
High-risk missing literature:
Opportunity hint:
Folder written:
Compact checklist status:

Reference Files

Load only what is needed:

• references/search-and-scoring.md: Use for source policy, source-quality exclusions, source tiers, paper-type taxonomy, and scoring anchors.
• references/report-template.md: Use when writing the literature-search folder files.

Retrieval Execution

Batch independent public-safe query clusters when supported, then deduplicate by DOI/arXiv identifier and normalized title before deeper reading. Verify important claims in the actual paper or primary page, not snippets. Track inspected sections and publication/version status. Do not score an unread method or treat inaccessible results as absent. An unavailable source limits that claim; continue with available primary evidence and report the relevant coverage gap.

For file outputs, follow ../ccf-common/references/artifact-contracts.md: resolve existing project paths first, keep generated working files under one stable task/artifact directory, and update canonical files in place. Load this shared policy only when files are written and it is not already in context.