AgentSkillsCN

paper-to-repro-spec

当被要求重现学术论文时，可将方法细节、超参数、评估协议以及图表定义提取至docs/repro_spec.md文件中。切勿随意猜测缺失细节；应列出所有模糊之处，并提出最简明的澄清实验方案。

SKILL.md
--- frontmatter
name: paper-to-repro-spec
description: Use when asked to reproduce an academic paper. Extract method details, hyperparameters, evaluation protocol, and figure definitions into docs/repro_spec.md. Do not guess missing details; list ambiguities and propose minimal disambiguation experiments.

You are reproducing a research paper.

Steps:

  1. •Read the paper PDF and any supplement in the repo.
  2. •Produce docs/repro_spec.md with:
    • •Task/environment definition (state, action, reward, termination, episode length)
    • •Model architecture (layers, sizes, activations, normalization)
    • •Learning algorithm (losses, targets, update rules, schedules)
    • •Training protocol (steps, episodes, replay, exploration, warmup)
    • •Evaluation protocol (how often, deterministic vs stochastic, number of eval episodes)
    • •Metrics + how plotted (smoothing, bins, mean/median, CI/SEM)
    • •Exact hyperparameter table (include defaults)
    • •“Unknowns” section: missing details + 2–3 plausible options
    • •“Disambiguation plan”: minimal tests to choose among options
  3. •Add a checklist mapping each reproduced figure panel to:
    • •input logs required
    • •code paths producing it
    • •acceptance criteria vs paper
  4. •Do not implement code until docs/repro_spec.md exists.