DewPoint
Use the running DewPoint app to connect the source, build its reusable data model and dashboards, and query numbers live. The result is a workspace and dashboard the user can reopen in the app.
Run the bundled helper with python3 <skill-directory>/scripts/dewpoint.py <command> using the directory that contains this SKILL.md. The helper targets http://127.0.0.1:8000 by default. For a team deployment, use its approved HTTPS API origin with --api ORIGIN before the command, or set DEWPOINT_API_URL. In a Claude desktop session, check connectivity from the skill runtime with status; localhost there may differ from localhost in the user’s browser. status should report database: true and cube: true. For a fresh stack, follow the repository README's Setup section and check that the backend has a configured LLM provider. The local dashboard UI defaults to http://localhost:5173.
Select the source
- Start with the decision, intended audience, and business question. Resolve ambiguities that change the answer, such as completed versus placed orders, unique orders versus line items, or the reporting period, time zone, and currency.
- If a matching workspace exists, use its ID from
workspaces. Reuse it when its source and model cover the question. Leave saved dashboards in place. - For a new source, have the user enter a read-only connection in DewPoint or supply its URL privately through
DEWPOINT_SOURCE_URLin the helper process environment. Thetablesandcreatecommands can also prompt interactively on a terminal. Keep connection URLs containing credentials out of messages, command arguments, source control, and draft files. For a local SQLite file, pass--sqlite /absolute/path.dbto both commands. The file must be accessible to the backend. - Run
tablesand identify the tables or views needed for the decision. When their purpose is unclear, use the DewPoint connection picker's table suggestions or ask for the relevant schema or views. Create the workspace withcreate --name NAME --goal GOAL --table SCHEMA.TABLE, repeating--tablefor each relevant table. Use the exact table name without a schema when the listed schema is null. Include confirmed definitions and date scope in the goal. New workspaces persist connection metadata and can send data samples to the stack's configured model provider. Use the data and provider approved for the task. - Record the returned workspace ID first, then use
wait WORKSPACE. If setup is still running, continue polling the same workspace. If it fails, use its status detail to resolve the issue and retry setup from DewPoint. If a create request disconnects, checkworkspacesbefore starting another.
Build and verify
- Run
catalog WORKSPACEanddashboards WORKSPACE. Review measure definitions, filters, table grain, and join relationships. DewPoint usually saves a starter dashboard during workspace setup. - If a saved dashboard already answers the question, run
verify WORKSPACE --template DASHBOARD_ID. Otherwise usedraft WORKSPACE --prompt 'Build and name a dashboard for …' > DRAFT.json. Include the approved definitions and request a layout that makes the decision easy to make. To start from a saved layout, add--template DASHBOARD_ID. A later draft can take--draft PREVIOUS.jsonand a new prompt. Use an agent-owned temporary directory for these files. The Designer can change the shared model through its tools. Request model changes only when the task calls for them. - Run
verify WORKSPACE --draft DRAFT.jsonto query every chart through Cube with the saved defaults. Check the numbers against the question and the catalog definitions. For join-sensitive totals, check the aggregation at the source grain or compare against an authorized read-only control query. Disclose unresolved definitions, missing values, incomplete coverage, and chart errors where they affect the answer. A successful query establishes that the chart runs. Use the business definition and a control total to establish that it answers the question. - When the user has asked for a dashboard and the verified draft answers the question, run
save WORKSPACE --draft DRAFT.json. This saves a new dashboard through DewPoint after verifying that every chart can load. If saving times out, checkdashboards WORKSPACEbefore trying again. - Return the app URL, workspace and dashboard names, the decision-relevant result, and the definition or check behind the main metric. Indicate whether the result is already saved or still a draft.