编写面向业务负责人的数据质量报告
依据已有检测结果,按关键字段呈现质量指标、业务影响和问题趋势,并明确整改责任与后续行动。
完整提示词
<context>
A data quality report is written for the people who can fix the data, usually the business owners of the process that enters it, not only for engineers. It is useful when it measures quality along clear dimensions and ties each problem to a consequence. Completeness: missing values, including disguised ones such as "N/A", "unknown", "0000", 1900-01-01 or a default country. Validity: values in the right format, range and list. Uniqueness: duplicate records, often the same customer entered twice with spelling differences. Consistency: the same fact agreeing across systems. Timeliness: data arriving within its agreed delay. Accuracy, checked against a trusted reference or a sample, is the hardest and is often not measured; the report says so. A single overall score hides the problem: 98% complete overall can mean the one field billing depends on is 70% complete. So results are reported per critical field, against a threshold, with the trend, and the root cause usually sits in an entry screen, an integration or a missing validation rule, not in the data team.
</context>
<task>
<profile_results>
{{profile_results}}
</profile_results>
{{#dataset_context}}
<dataset_context>
{{dataset_context}}
</dataset_context>
{{/dataset_context}}
1. State the scope: dataset, period, record count, which checks were run, and which dimensions were not measured.
2. Classify each field as critical or not, using the context given (what it feeds). If no context is given, propose a classification and label it "proposed".
3. Build the scorecard per critical field and dimension: the measured rate, the threshold (agreed if given, otherwise a proposed one labelled as such), status, and change since the last period. Count disguised missing values as missing and say which ones you counted.
4. List issues ranked by business impact: what is wrong, how many records, what it breaks downstream, the likely root cause (labelled "likely" unless confirmed), and the fix (at source, in the integration, or a cleanup).
5. Describe trends: what is improving, what is worsening, and any step change with its date.
6. Assign each issue to an owner role (or the named owner if given), with a next action and a review date.
7. State the limits: dimensions not measured, sampling, and checks that could give false passes.
</task>
<constraints>
- Use only the measurements given; do not invent counts, thresholds presented as agreed, owners or root causes.
- Report rates with the record counts behind them.
- Do not roll everything into one overall quality score; if one is requested, show it alongside the per-field results.
- Do not include actual personal data values from the records in the report; describe patterns instead.
- Keep the tone factual and blame-free; describe processes, not people.
</constraints>
<output_format>
## Summary
Three bullets: overall state, the worst issue and its impact, the main action.
## Scope and method
## Scorecard
Table: field | dimension | rate | records affected | threshold | status | change.
## Issues by impact
Table: issue | records | downstream impact | likely root cause | fix.
## Trends
## Owners and actions
Table: issue | owner | next action | review date.
## Limits
</output_format>可替换内容
- {{profile_results}}
- {{#dataset_context}}
- {{dataset_context}}
- {{/dataset_context}}
分类与标签
办公效率、数据质量报告、关键字段检查、业务整改、责任分配