AgentSkillsCN

fill-results

根据输出文件中的最新实验结果，更新论文中的表格。当实验完成、结果需要纳入论文时，可使用此功能。

SKILL.md
--- frontmatter
name: fill-results
description: Update paper tables with new experimental results from output files. Use when experiments complete and results need to flow into the paper.
disable-model-invocation: true
allowed-tools: Read, Write, Edit, Grep, Glob, Bash

Fill Results Into Paper Tables

Update paper tables with new experimental results. Argument: path to results file or directory.

Source: $ARGUMENTS

Step 1: Understand the Results

Read the results file(s). Determine the format:

  • •JSON with metrics per dataset/config
  • •CSV/TSV with rows of results
  • •Log files with metrics printed at end
  • •Python pickle/numpy (extract via script)

Extract all metric values and their associated conditions (dataset, model variant, budget level, etc.).

Step 2: Map Results to Tables

For each table in the paper:

  1. •Read the table structure (columns, rows, what each cell represents)
  2. •Match result metrics to table cells by:
    • •Dataset name
    • •Model/method variant
    • •Metric name (accuracy, F1, exact match, etc.)
    • •Budget/intervention level
  3. •Identify which cells have new values vs. which should keep existing values

Step 3: Update Tables

For each cell that has a new value:

  1. •Compare with existing value (if any)
  2. •If the cell currently has a placeholder (---, TBD, ???), replace with new value
  3. •If the cell has an old value, replace it
  4. •Match the decimal precision of surrounding cells
  5. •Update any bold/underline formatting:
    • •Bold = best in column (recompute after updating)
    • •Underline = second-best (recompute after updating)

Step 4: Update Prose

Search for prose that references updated numbers:

  1. •Find paragraphs near \ref{tab:...} for each updated table
  2. •Check if any numbers in prose now need updating
  3. •Check if superlatives ("best", "highest", "outperforms") are still correct
  4. •Check if improvement claims ("+3.2%", "by a margin of") need recalculation

IMPORTANT: Do not silently change prose numbers. For each prose number that needs updating, show the old and new value and ask for confirmation before editing.

Step 5: Verification

After all updates:

  1. •Compile the paper (invoke compile-latex skill if available)
  2. •Run check-tables to verify consistency
  3. •Report summary:
code
RESULTS UPDATE SUMMARY
======================
Source: [results file path]
Tables updated: N
Cells updated: N (of which N were placeholders, N were value changes)

Changes:
  tab:main_results — 12 cells updated, 3 bold markers moved
  tab:ablation — 4 cells updated (were TBD)
  
Prose updates needed:
  Line 142: "82.3%" → should be "84.1%" (awaiting confirmation)
  Line 148: "improves by 3.2%" → should be "improves by 4.5%"

Remaining placeholders: N cells still have TBD/--- values