Applies to SaaS, product growth, conversion funnels, and retention; when users have only transaction data, prioritize order and revenue analysis, and do not fabricate active user or subscription metrics.
Execution Points
Use a light gray-white canvas, white panels, a blue primary series, and a purple comparison series. Set one main trend area per page, with 3–4 compact KPIs followed by trends, funnels, or retention tables. Green indicates only verified positive change; a decline in paid cancellation rate and a decline in revenue cannot be judged as good or bad by the same rule. Prioritize calculating current-period/prior-period windows of equal length, and note whether the period is complete or incomplete. Define retention cohorts by first-occurrence event, deduplicate with a stable user key, and show periods not yet reached as unobserved rather than 0. Funnel stages require the same entity, time window, and ordered conditions; when event records are missing, do not mislabel independent counts as conversions. Primary color #6461E8, comparison #8B79CF, page #F3F5FB, text #242C46, and panels keep clear boundaries.
Conversation Examples
When asked "How has growth been recently," first read the existing dates and metrics, and answer with trends and contribution breakdowns; if there are no user events, explicitly use order/customer trends as a substitute and explain the data limitations.
How DB-GPT Works
- Communicate in the user's language. First infer intent from the selected data, prior context, dashboards, and annotations; when the expression is vague but actionable, state one reasonable assumption and continue. Ask only one specific question for gaps that would change the result, and do not repeat generic clarification phrases.
- The selected database is for read-only discovery only: confirm fields, granularity, units, date range, null values, and unique keys. When there is no data, layout can still be discussed; before generating actual metrics, ask the user to select a data source. Illustrative values must be clearly labeled and never disguised as query results.
- When creating an editable dashboard, call the existing dashboard tool and load the dashboard-builder workflow: first plan_dashboard, then create according to the existing confirmation process; when modifying an existing dashboard, first load_dashboard_draft, then propose changes matching the version. The visual specification supplements the business plan and cannot skip SQL validation, saving, or publishing rules.
- Use the report generation tool only when the user explicitly requests an HTML report; a native dashboard must not be secretly replaced with an image or standalone HTML. The native schema uses only the fields, chart types, and theme properties allowed by the tool, and does not pass arbitrary CSS or nonexistent layout enums. Decoration that cannot be fully expressed should be reduced to usable color schemes/layouts, with differences explained.
- Before editing, preserve the user's data, metric definitions, and unmodified components. Metric changes must update titles, units, and calculation definitions in sync; for ratio aggregation, sum numerators and denominators first. Do not fabricate targets, year-over-year comparisons, or forecasts when there is no target data.
- Before delivery, check empty results, long labels, narrow screens, light and dark themes, legends, zero values, and missing values; animations should respect the user's reduced-motion preference. Report only execution results confirmed by the tool.
Resources and Boundaries
This directory is a DB-GPT product skill: it does not modify the host's global configuration, does not install external scripts, and does not send data to reference websites. Use fonts already provided by the system and project; when external fonts are missing, use Chinese fallbacks. The interface shows a style demonstration; business values come from the selected data in actual use.
Reference Sources
Reference Sive's related skills scenario and visual direction; the rules have been rewritten by DB-GPT Dashboard Lab and are not an official Sive release package.