菜单
Use when the user is new to a research direction and wants to scan papers to understand branches, representative works, common methods, benchmarks, metrics, saturated areas, and unclear areas.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when the user knows the broad field and reads papers to find concrete, researchable, experimentally testable gaps or research questions.
Use when the user wants to stress-test whether one selected Research Question Card's problem is real, evidence-backed, externally grounded, safely motivated, and worth keeping as a research motivation.
Use when the user has a research problem but needs theoretical foundations, conceptual lineage, or research traditions that support the problem framing.
Use when the user has a rough problem and reads papers to collect transferable modeling ideas, architectures, training flows, optimization objectives, agent pipelines, or data construction methods.
Use when the user has a selected Candidate Method, multiple Candidate Methods to merge, a user-authored rough method, or a reconstructed chat method and wants to converge on one human-owned method outcome before experiment design, research framing, or risk review.
Use when the user has a method idea or Committed Method Design and reads papers to design tasks, datasets, baselines, metrics, ablations, human evaluation, user studies, or a claim-to-evidence experiment plan.
| name | paper-reading-field-map |
| description | Use when the user is new to a research direction and wants to scan papers to understand branches, representative works, common methods, benchmarks, metrics, saturated areas, and unclear areas. |
Use this skill to run a complete Field Map Workflow: define the field boundary, collect seed papers, expand the literature set, extract paper position records, cluster the field, identify research opportunity candidates, and produce a static HTML report.
The goal is not to fully summarize every paper. The goal is to build a durable map of the field: what problems exist, what method routes exist, which assumptions they rely on, how they are evaluated, where papers agree or conflict, and what gaps remain.
Set {workspace-root} before creating, scanning, or updating artifacts:
{workspace-root} to workspace (the repo-local workspace/ directory).workspace/ layer.{workspace-root} from that path and keep related artifacts under the same root.{workspace-root}/field-maps/{field-slug}/.field_boundary.md.seed_papers/.search_strategy.md, including candidate discovery directions and selection criteria.all_papers/.research_clusters.md.research_opportunities.md.index.html report that links to the markdown artifacts.Do not skip these gates:
If the user gives corrections at any gate, revise the artifact or proposal and ask for confirmation again.
Start by drafting a Proposed Field Boundary instead of asking the user to fill out a blank questionnaire. The proposed boundary must include:
Ask the user what is inaccurate, missing, or too broad. Continue revising until they confirm. Then write the confirmed boundary to field_boundary.md using references/field-boundary-template.md.
If the user directly uploads or links papers, prioritize those User-Provided Papers.
If the user does not directly provide papers, use Network Discovery by default. Search online sources such as survey bibliographies, citation trails, benchmark pages, conference proceedings, leaderboards, and academic search engines.
For every paper, record:
If network access is unavailable, do not pretend discovery is complete. Tell the user what could not be searched and ask for paper links or permission to continue when network access is available.
The Seed Set should contain roughly 8-12 papers. Prefer coverage over exact count:
If a category is absent in the field, record that absence in search_strategy.md; it may indicate an Evaluation Gap or System Gap.
Before reading seed papers, present the proposed Seed Set with a short reason for each paper and wait for confirmation.
Use Position-Oriented Reading. Read only deeply enough to complete the Paper Position Record and place the paper in the field.
Default reading order:
Read methods, formulas, appendices, or experiments more deeply only when the paper is foundational, defines a benchmark, or determines a cluster boundary.
After seed paper records are complete, draft search_strategy.md before reading expansion papers. The strategy must include:
Use a Candidate Pool first, then select roughly 25-35 papers as the Selected Expansion Set. Create one Paper Position Record per selected paper in all_papers/.
Cluster all read papers into no more than 10 Research Clusters.
Each paper must have exactly one Primary Research Cluster and may have zero or more Secondary Research Clusters. If more than 10 clusters appear necessary, merge them into higher-level clusters and record the boundary tradeoff.
Do not list unsupported speculation as a Research Opportunity Candidate. Each opportunity must be backed by links to specific paper records and must be phrased so it can become a possible research question.
Use these gap types:
Create a zero-build static index.html inside the Field Map Workspace. It should be readable by opening the file directly and should link to the markdown artifacts.
Do not introduce a frontend framework, build step, or external runtime for the report.
Use the reference templates in this directory:
references/workspace-structure.mdreferences/field-boundary-template.mdreferences/paper-position-record-template.mdreferences/search-strategy-template.mdreferences/research-clusters-template.mdreferences/research-opportunities-template.mdreferences/html-report-template.mdStop when the user can explain: