菜单
On-demand time-series forecasting. CAPTURE params from project context, call run_forecast, present deterministic engine results.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Use when helping initialize, configure, or prepare a Seeknal project like a coding agent
Translate business questions into metrics, SQL evidence, and actionable recommendations
Run multi-step SQL plus Python/statistics/ML analysis while keeping tools thin and evidence grounded
Answer business questions from read-only connected databases using deterministic schema discovery and SQL evidence
Run Python code in an isolated subprocess for statistical/ML/visualization work beyond what SQL can express
End-to-end workflow for adding a new pipeline node to a seeknal project — scaffold, validate, apply, and (optionally) run via the 5 thin pipeline-build tools
| name | forecasting |
| description | On-demand time-series forecasting. CAPTURE params from project context, call run_forecast, present deterministic engine results. |
| tags | ["forecasting","time-series","prediction"] |
| version | 1.0.0 |
Read project context (context/forecast_guide.md) to determine:
Then construct a 2-column SQL (X = time grain, Y = value):
SELECT date_trunc('month', <date_col>) AS x, COUNT(DISTINCT <id_col>) AS y
FROM <schema>.<table>
WHERE <date_col> >= '<baseline>'
GROUP BY 1
ORDER BY 1
Lock: {sql, periods}. If the source/series is ambiguous → call request_clarification.
Call run_forecast(sql, periods) with the SQL you constructed and the
horizon in periods. The tool executes the SQL, validates it is exactly
2 columns with enough rows, infers the frequency, and triggers the engine.
Do NOT do forecast arithmetic.
If the tool returns ## Ditolak → present the reason to the user.
Do not retry with different params unless the user changes the request.
Format the tool's structured output for the user:
agent.forecast.enabled: true is set in seeknal_agent.yml.