ppw:experiment

Purpose

This Skill accepts experiment result data — tables, statistics, or result descriptions — and runs a two-phase workflow. Phase 1 extracts measurable findings from the data and presents a structured Finding list for user confirmation. Phase 2 generates discussion paragraphs for each confirmed finding, using grounded evidence language followed by calibrated interpretation. Literature connections are never invented: the Skill asks the user to provide prior work, and writes [CONNECT TO: ...] placeholders when none is supplied. The Skill serves researchers preparing results and discussion sections for journal or conference submission.

Core Prompt

Source: awesome-ai-research-writing — 实验分析

# Role
你是一位具有敏锐洞察力的资深数据科学家，擅长处理复杂的实验数据并撰写高质量的学术分析报告。

# Task
请仔细阅读我提供的【实验数据】从中挖掘关键特征、趋势和对比结论，并将其整理为符合顶级会议标准的 LaTeX 分析段落。

# Constraints
1. 数据真实性：
   - 所有结论必须严格基于输入的数据。严禁编造数据、夸大提升幅度或捏造不存在的实验现象。
   - 如果数据中没有明显的优势或趋势，请如实描述，不要强行总结所谓的显著提升。

2. 分析深度：
   - 拒绝简单的报账式描述（例如不要只说 A 是 0.5，B 是 0.6），重点在于比较和趋势分析。
   - 关注点包括：方法的有效性（SOTA 比较）、参数的敏感性、性能与效率的权衡，以及消融实验中的关键模块贡献。

3. 排版与格式规范：
   - 严禁使用加粗或斜体：正文中不要使用 \textbf 或 \emph，依靠文字逻辑来表达重点。
   - 结构强制：必须使用 \paragraph{核心结论} + 分析文本 的形式。
     * \paragraph{} 中填写高度凝练的短语结论（使用 Title Case 格式）。
     * 紧接着在同一段落中展开具体的数值分析和逻辑推演。
   - 不要使用列表环境，保持纯文本段落。

4. 输出格式：
   - Part 1 [LaTeX]：只输出分析后的 LaTeX 代码。
     * 必须对特殊字符进行转义（例如：`%`、`_`、`&`）。
     * 保持数学公式原样（保留 `$` 符号）。
     * 不同的结论点之间请空一行。
   - Part 2 [Translation]：对应的中文直译（用于核对数据结论是否准确）。
   - 除以上两部分外，不要输出任何多余的对话。

Trigger

Activates when the user asks to:

• Analyze experiment results, identify patterns, or extract findings from result data
• Generate discussion paragraphs from confirmed findings
• 分析实验结果、识别规律、生成讨论段落

Example invocations:

• "Analyze my results table and write discussion"
• "帮我分析实验结果并写讨论段"
• "Generate discussion paragraphs for my findings"
• "What patterns do my experiment results show?"

Modes

Default mode: direct. User provides result data and gets Phase 1 finding list, confirms, then receives Phase 2 discussion paragraphs.

Mode inference: "Just identify findings" or "只分析不写讨论" runs Phase 1 only.

References

Required (always loaded)

Leaf Hints (loaded in Phase 2)

Conditional

Ask Strategy

Before starting, ask about:

1. Research questions: "What are the main research questions this experiment addresses?" (Required — Phase 2 uses these to connect findings to purpose)
2. Prior work to connect to: "Which papers or findings should the discussion reference?" (Optional — ask once; if declined, use [CONNECT TO: ...] placeholders in Phase 2)
3. Target journal (if not specified): ask once; if declined, use general academic style

Rules:

• Never ask more than 3 questions before starting Phase 1
• Research questions are mandatory; the Skill cannot produce grounded Phase 2 output without them
• If the user declines to provide research questions, write [RESEARCH QUESTION: describe your RQ here] placeholders rather than blocking the workflow entirely

Workflow

Step 0: Workflow Memory Check

• Read .planning/workflow-memory.json. If file missing or empty, skip to Phase 1.
• Check if the last 1-2 log entries form a recognized pattern with ppw:experiment that has appeared >= threshold times in the log. See skill-conventions.md > Workflow Memory > Pattern Detection for the full algorithm.
• If a pattern is found, present recommendation via AskUserQuestion:
  • Question: "检测到常用流程：[pattern]（已出现 N 次）。是否直接以 direct 模式运行 ppw:experiment？"
  • Options: "Yes, proceed" / "No, continue normally"
• If user accepts: set mode to direct, skip Ask Strategy questions.
• If user declines or AskUserQuestion unavailable: continue in normal mode.

Phase 1: Analyze Results

Step 1 — Prepare:

• Load references/expression-patterns.md overview
• If a journal was specified, load its template; if template is missing, refuse with message above
• Read input: file via Read tool, pasted results block (table, statistics, narrative), or structured_data
• Opt-out check: Scan the user's trigger prompt for any of these phrases (case-insensitive, exact phrase match): english only, no bilingual, only english, 不要中文. Store result as bilingual_mode (true/false). This flag governs Phase 2 bilingual output below.
• Guard — measurable data required: if input is vague (e.g., "my results show improvement" without values, comparisons, or metrics), refuse: "Please provide specific values, comparisons, or metrics before I can identify findings."
• LaTeX table input: read data values and captions; ignore typesetting commands
• Record workflow: Append {"skill": "ppw:experiment", "ts": "<ISO timestamp>"} to .planning/workflow-memory.json. Create file as [] if missing. Drop oldest entry if log length >= 50.

Step 2 — Extract Findings:

• Identify measurable comparisons: method A vs. method B, magnitude, direction
• Identify trends: performance across conditions, dataset sizes, subgroups
• Identify outliers: results that deviate from the overall pattern
• Each finding must include: a direction (higher/lower/better/worse), a magnitude or value, and a comparison group or condition

Step 3 — Present Finding List:

• Use locked format per item:
  Copy

  Finding 1: [subject] [comparison/trend] [value] on [metric/condition]
  Finding 2: Performance degrades in [condition] ([N] vs. [M])
  Finding 3: [Subgroup] shows the largest effect ([value])
  

• Summary line: "Identified N findings. Please confirm, correct, or add before I write discussion."
• Wait for user approval before proceeding to Phase 2

Phase 2: Generate Discussion

Step 1 — Prepare:

• Load references/expression-patterns/results-and-discussion.md for evidence reporting language
• Load references/expression-patterns/conclusions-and-claims.md for calibrated interpretation
• Load references/anti-ai-patterns/vocabulary.md to screen output before presenting
• Hold any user-provided prior work for connection sentences

Step 2 — Write Discussion Paragraphs:

• Follow the Core Prompt constraints above as the primary instruction set for analysis and output formatting.
• One paragraph per confirmed finding
• Each paragraph follows this structure:
  1. Evidence sentence: state the finding with full quantification (use results-and-discussion.md patterns for comparative and trend language)
  2. Interpretation sentence: claim using calibrated language from conclusions-and-claims.md ("suggests", "indicates" — never lead with interpretation before evidence)
  3. Connection sentence: if user provided prior work, connect the finding to it; otherwise write [CONNECT TO: describe the prior finding here]
• CRITICAL rule: Interpretation sentence must follow the evidence sentence. Never open a paragraph with an interpretive claim without first stating the quantified evidence.
• After generating all paragraphs, check output against vocabulary.md; revise any flagged patterns

Step 3 — Output:

• Present all discussion paragraphs in sequence

• Bilingual display: If bilingual_mode is true: after each discussion paragraph, append a > **[Chinese]** ... blockquote containing the Chinese translation of that paragraph. Use a section header "双语对照 / Bilingual Comparison:" before the first paragraph. Format per finding paragraph:

  [English discussion paragraph for Finding N]

  [Chinese] [Chinese translation of the discussion paragraph for Finding N]

• Do not insert Chinese into any written file. If the user requested writing discussion to the paper file via Write tool, write English-only paragraphs to the file; the Chinese blockquotes remain in conversation only.

• If bilingual_mode is false (opt-out detected): skip bilingual display entirely.

• If file input was used, offer to append discussion to file using Write tool

• Recommend Polish Skill for further expression refinement if higher-register prose is desired

Output Contract

Note: Phase 2 output cannot be produced without Phase 1 confirmation. If user skips Phase 1 and requests discussion directly, require Phase 1 completion first.

Edge Cases

Fallbacks

Examples

Minimal invocation: User pastes a results table comparing Method A and Method B on accuracy and F1 score. User states RQ: "Does our approach outperform the baseline on both metrics?"

Phase 1 output:

Finding 1: Method A outperforms Method B by 3.2 percentage points on accuracy (87.4% vs. 84.2%)
Finding 2: Method A outperforms Method B by 4.1 points on F1 score (82.6 vs. 78.5)

Identified 2 findings. Please confirm, correct, or add before I write discussion.

User confirms. No prior work provided.

Phase 2 output (Finding 1):

Method A achieves 87.4% accuracy, outperforming Method B by 3.2 percentage points (84.2%).
This suggests that the proposed approach captures more discriminative features for the task,
yielding a consistent accuracy gain across evaluation conditions.
[CONNECT TO: describe a prior finding showing similar accuracy improvements for this approach]

Skill: experiment-skill Conventions: references/skill-conventions.md