zouchenzhen/baoyan-program-research-skill/baoyan-program-research

Baoyan Program Research

Use this skill to investigate 推免/保研 opportunities for any major. The output should be an evidence-backed, privacy-safe candidate list, usually a CSV, not a prose-only recommendation.

Core Rules

• Treat official notices as the source of truth. Aggregators, forums, WeChat reposts, and prior spreadsheets are leads only.
• Do not invent dates, quotas, employment outcomes, salary, or eligibility gates. Use 待核验 / 官网未公开 when evidence is missing.
• Search all common names: 夏令营, 暑期学校, 科学营, 优秀大学生, 优材生, 预推免, 预选拔, 推免复试, 推荐免试, 九推, 接收推免生.
• Build a broad pool first, then rank it. Include strong universities, relevant colleges, CAS or other research institutes, industry institutes with master's authority, and practical employment-oriented options.
• Preserve privacy. Never include personal phone numbers, email addresses, ID numbers, student IDs, full names from prior-destination files, private local paths, or raw resumes in a public artifact.
• Separate three labels in the CSV: official facts, inferred fit, and local historical-success evidence.

Inputs To Collect

Ask for or infer, when available:

• Candidate profile: school tier, major, rank/GPA, English scores, research/project keywords, awards, desired degree type, city preference, employment vs PhD orientation.
• Local files: resume, transcript summary, award list, prior-destination spreadsheets, old application lists.
• Target policy: broad sea application vs curated list, master's only vs master/PhD, employment-oriented vs academic.
• Output path and format. Default to CSV plus a short Markdown summary.

Workflow

1. Inventory and redact inputs

• List local files first; inspect size before reading.
• Extract only decision-relevant facts from resumes.
• Do not copy personal contact details into working notes or public repos.

1. Profile fit model

• Derive discipline keywords in Chinese and English.
• Identify adjacent departments that can admit the major, such as chemistry for polymers, EE for electronic materials, public health for biomedical engineering, or management science for data/business majors.
• Record hard gates: rank cutoff, English cutoff, required recommendation letters, direct-PhD-only notices, master's-friendly notices.

1. Coverage plan

• For broad tasks, explicitly include national coverage checks: 985, strong 211/double-first-class programs, top discipline-specific schools, CAS/institute systems when relevant, and local historical-success destinations.
• For employment-oriented candidates, score city and industry resources separately from academic prestige.
• For PhD-averse candidates, mark direct-PhD-heavy programs as risk even when the school is strong.

1. Search and verification

• Use current-year official pages first. If current-year pages are not published, use the latest three years as timing references.
• Search each target with multiple notice names and department names.
• Preserve original URLs and exact dates where available.
• If a page is image/PDF-heavy, extract enough text to verify dates and eligibility, then mark any uncertain fields.

1. Prior-destination evidence

• Aggregate local prior-destination files by destination school/institute and college/department.
• Export only counts and summaries, never student names.
• Match universities by exact school name. Match institutes by specific institute tokens. Do not match on generic labels like 材料学院, 计算机学院, or 化学学院 alone.
• Treat prior success as a priority booster, not proof of admission probability.

1. CSV delivery

• Use the schema in references/output_schema.md.
• Include official status, dates, links, source quality, local prior-success count, fit notes, risks, and action recommendation.
• Add a coverage summary: row count, 985 coverage if applicable, institute count if applicable, current-year notices found, and prior-success hit count.

1. Quality gate

• Run scripts/validate_candidate_csv.py on generated CSVs when possible.
• Run scripts/check_privacy.py before publishing or sharing any repo/artifact.
• Re-open the CSV summary with a limited preview to catch encoding, delimiter, and field-shift issues.

When To Load References

• Need exact CSV columns: read references/output_schema.md.
• Need search query patterns and source ranking: read references/search_strategy.md.
• Need privacy handling rules: read references/privacy_and_safety.md.
• Need local prior-destination aggregation guidance: read references/prior_destination_matching.md.

Helper Scripts

• scripts/aggregate_prior_destinations.py: aggregate .xlsx, .xls, and .csv prior-destination files into privacy-safe counts.
• scripts/validate_candidate_csv.py: check required CSV columns and obvious source/date issues.
• scripts/check_privacy.py: scan files for common PII, local paths, tokens, and private source names before release.

Final Response Pattern

Report:

• output file path,
• coverage numbers,
• current-year urgent notices,
• strongest local prior-success destinations,
• unresolved verification gaps,
• privacy status.

Keep direct recommendations short and grounded in the CSV fields.