Survey analytics

Survey Analytics

This skill analyzes survey results.

It does not generate survey HTML. It consumes:

• a survey schema
• a survey answers dataset

Primary job

Take a validated survey schema plus answers and produce:

• one Excel workbook where columns are questions and rows are submissions
• per-question statistics and charts by question type
• cross-question or logic-driven analysis tables
• a machine-readable analysis JSON
• a human-readable insight report

Read these first

Before analyzing, read:

• ../../specs/analytics-input.md
• ../../specs/answer-storage.md
• ../../specs/submission-api.md
• ./references/metrics-model.md
• ./references/report-shape.md
• ./docs/OUTPUTS.md

Expected input

Preferred input:

• absolute path to a survey schema JSON file
• absolute path to an answers JSON or JSONL file

Also acceptable:

• inline schema JSON
• inline answers JSON
• a directory containing both schema and answers exports

Required execution mode

Do not stop at a prose-only analysis when local file execution is available.

Use the executable pipeline:

python3 <survey-analytics-root>/tools/analyze_survey_results.py \
  --schema /absolute/path/to/schema.json \
  --answers /absolute/path/to/answers.json \
  --out-dir /absolute/path/to/output-dir

When the user wants specific cross-question analysis, pass a config file:

python3 <survey-analytics-root>/tools/analyze_survey_results.py \
  --schema /absolute/path/to/schema.json \
  --answers /absolute/path/to/answers.json \
  --cross-config /absolute/path/to/cross-config.json \
  --out-dir /absolute/path/to/output-dir

The script generates:

• <name>.analysis.xlsx
• <name>.analysis.json
• <name>.analysis.md

If the user only asks for insights and does not want files, you may summarize the generated report instead of returning file contents.

Analysis rules

1. validate that every normalized answer document belongs to the provided surveyId
2. ignore malformed records, but report how many were excluded and why
3. compute distributions only from structurally valid answers
4. preserve question ids and option ids in intermediate reasoning
5. separate factual summary from interpretation
6. call out low sample size before making strong conclusions
7. do not invent respondent intent that is not grounded in the dataset
8. when schema logic exists, use it to infer finish-path or segment-level analysis when possible

Output requirements

The workbook should include at least:

• Overview
• Responses
• Question Summary
• Cross Tabs
• Finish Sentiment
• Extra Analysis
• Insights
• one dedicated sheet per question for chartable output

The flattened response table should use:

• fixed metadata columns first
• then one column per question
• cell values should be respondent-readable, not raw nested JSON unless unavoidable

Chart rules by question type

• radio: option distribution column chart
• checkbox: option selection frequency column chart
• input: no forced chart; keep a readable sample table instead
• score: score distribution chart and average score
• nps: score distribution chart plus promoter/passive/detractor band summary

Text analysis rule

For free-text fields, the skill should distinguish between:

• feedback/opinion text
• information collection text

Only feedback/opinion text should receive:

• theme extraction
• sentiment analysis using business-friendly Chinese labels such as:
  • 高度满意
  • 满意
  • 中性
  • 不满
  • 强烈不满

Do not run sentiment analysis on fields that are clearly for information collection, such as:

• name
• phone
• email
• company
• contact info
• address

These fields may still appear in the flattened workbook, but they should not drive sentiment findings.

Cross analysis strategy

Default cross analysis should prefer:

• radio vs score
• radio vs nps
• radio vs checkbox
• finish-path distribution when logic contains end_survey
• extra attribution dimensions such as utm_source / campaign vs finish-path and text sentiment
• logic-driven path summaries for branch-heavy surveys

If the user explicitly names segment questions or wants tighter control, create a cross config JSON with this shape:

{
  "pairs": [
    {
      "segmentQuestionId": "radio-205152",
      "metricQuestionId": "score-205160"
    }
  ]
}

Only those pairs should be generated when a cross config is provided.

When the user wants explicit control over which text fields receive theme/sentiment analysis, pass:

python3 <survey-analytics-root>/tools/analyze_survey_results.py \
  --schema /absolute/path/to/schema.json \
  --answers /absolute/path/to/answers.json \
  --analysis-config /absolute/path/to/analysis-config.json \
  --out-dir /absolute/path/to/output-dir

Example:

{
  "textAnalysis": {
    "includeQuestionIds": ["checkbox-205170"],
    "excludeQuestionIds": ["input-205179"],
    "includeLabels": ["补充说明", "其他"],
    "excludeLabels": ["姓名", "手机号", "联系方式"]
  }
}

Rules:

• if a text field matches excludeQuestionIds or excludeLabels, do not analyze its sentiment
• if includeQuestionIds or includeLabels matches, do analyze it
• when no config is provided, fall back to the built-in heuristic

Minimum output

Return:

1. workbook path
2. analysis JSON path
3. insight markdown path
4. short summary of what was generated
5. any data-quality warnings

Output style

Keep the analysis decision-useful.

Do not dump raw counts without interpretation. Do not make causal claims unless the data supports them.