AgentSkillsCN

uv-create-paper-exercises

根据研究论文（arXiv URL/PDF）创建学习练习。当您希望将论文转化为：(1) 编程练习（将核心技术提炼为带测试的编程问题），以及 (2) 模型化练习（将公式与推理转化为带详细解答的计算问题）。在 $HUMAN_MATERIAL_PATH/exercises/<paper_slug>/ 下生成练习包，其中包含本地迷你技能，用于检验答案并揭示黄金解法。

SKILL.md
--- frontmatter
name: uv-create-paper-exercises
description: "Create learning exercises from a research paper (arXiv URL/PDF). Use when turning a paper into (1) a programming exercise (extract the core technique into a coding problem with tests) and (2) a modeling exercise (extract formulas/reasoning into calculation problems with worked solutions). Generates an exercise pack under $HUMAN_MATERIAL_PATH/exercises/<paper_slug>/ including local mini-skills to check answers and reveal golden solutions."

Create exercises from a paper

Goal

From a single paper, generate two exercise tracks:

  1. •Programming: turn the central technique into a coding problem with a crisp spec + tests.
  2. •Modeling: turn the paper’s formulas/derivations into calculation problems with worked solutions.

The output should be an exercise pack with:

  • •Clear prompts
  • •Golden solutions
  • •Automated checkers
  • •Local mini-skills to check answers on demand

Output location

Create the pack under:

  • •$HUMAN_MATERIAL_PATH/exercises/<paper_slug>/

Downloads/clones should go under (gitignored):

  • •$HUMAN_MATERIAL_PATH/.references/ (PDFs, arXiv zips, cloned repos)

Scaffold the pack

Use the scaffold script to create the file structure:

bash
python skills/uv-create-paper-exercises/scripts/scaffold_exercise_pack.py --slug <paper_slug>

Then fill in the prompts/solutions based on the paper.

Programming exercise workflow

  1. •Find the central technique:
    • •Identify the smallest “core loop” that makes the method work (data structure + algorithm).
  2. •Extract a clean problem:
    • •Define input/output precisely.
    • •Specify constraints.
    • •Provide 2–3 small examples.
  3. •Create a testable interface:
    • •Implement as a single function in my_solution.py (user) and solution.py (golden).
  4. •Write tests:
    • •Include edge cases and one randomized property check (if applicable).
  5. •(Optional) Use the paper’s code repo:
    • •If the paper links code, clone it under $HUMAN_MATERIAL_PATH/.references/repos/.
    • •Read it for correctness and for test cases; do not copy large chunks verbatim.

Modeling exercise workflow

  1. •Locate the modeling logic:
    • •The key formula(s), scaling law(s), or performance reasoning in the paper.
  2. •Turn it into 3–6 questions:
    • •Use numbers so the learner can compute concrete outputs.
    • •Include at least one “interpret the result” question.
  3. •Provide golden answers + explanations:
    • •answers.json for numeric targets
    • •SOLUTION.md for step-by-step reasoning
  4. •Add a checker:
    • •Compare my_answers.json against answers.json with tolerances.

Local answer-checking mini-skills

The scaffold creates two local mini-skills under:

  • •$HUMAN_MATERIAL_PATH/exercises/<paper_slug>/skills/uv-check-programming/
  • •$HUMAN_MATERIAL_PATH/exercises/<paper_slug>/skills/uv-check-modeling/

Use them to:

  • •Run checks
  • •Print diffs / failed tests
  • •Reveal golden solutions + explanations when asked