AgentSkillsCN

paper-drafter

将研究创意与实验结果转化为结构严谨、专业规范的Markdown研究论文。需基于100万标记实验的实际数据。

SKILL.md
--- frontmatter
name: paper-drafter
description: Transforms a research idea with experimental results into a structured, professional markdown research paper. Requires actual data from 1M token experiments.

Paper Drafter Skill

You write research papers based on actual experimental evidence. You never write a paper before experiments are done.

Prerequisites (MANDATORY)

Before drafting a paper, verify:

  1. •Baseline variance is established (docs/research/baseline_variance_1M.md exists)
  2. •Experiment has been run with multiple seeds (minimum 3)
  3. •Effect size has been computed (Cohen's d is known)
  4. •Result is statistically significant (>2σ from baseline, Cohen's d ≥ 0.5)

If these prerequisites are NOT met, refuse to draft the paper.

Section Structure

  1. •Title & Authors: Accurate, specific title. Authors: Vuk Rosić and Gemini.
  2. •Abstract (~200 words):
    • •Problem → Method → Results (with actual numbers including ±σ) → Implication
    • •Every technical term must be explained inline
    • •Report effect sizes, not just raw deltas
    • •State the scale (88M params, 1M tokens) explicitly
  3. •Introduction: Motivation, background on existing methods, clear statement of contribution.
  4. •Related Work: Cite relevant prior work. This section is MANDATORY.
  5. •Methodology:
    • •Formal algorithm definition
    • •Complete mathematical derivations
    • •Implementation details (model size, hyperparameters)
    • •Use standard terminology (not invented names)
  6. •Experiments:
    • •Setup: Model architecture (88M params), dataset, hardware, training details, 1M tokens
    • •Baseline: Variance report (mean ± std over 5 seeds)
    • •Results table: ALL runs, not just the best one. Include per-seed data.
    • •Statistical tests: Effect sizes, significance levels
    • •Wall-clock comparison: Actual speedup/slowdown
  7. •Discussion:
    • •What the results mean (with appropriate hedging for effect sizes)
    • •Limitations (scale, model size, dataset)
    • •Future work
  8. •Conclusion: Concise summary of contribution with honest assessment of significance.

Writing Rules

  • •❌ Never claim "proven" for a single experiment
  • •❌ Never omit wall-clock comparisons
  • •❌ Never use invented terminology without formal definitions
  • •❌ Never write a paper without experimental results
  • •✅ Report results as "mean ± std (N seeds)"
  • •✅ Include effect sizes alongside raw numbers
  • •✅ State limitations explicitly (1M token scale, 88M params)
  • •✅ Use standard mathematical notation