Synthefy

Use this skill any time a FACTUAL SUMMARY / PREDICTIVE SIGNALS text string is created, generated, modified, or used with Migas-1.5. This includes: calling generate_summary(), writing summaries manually, creating counterfactual scenarios with splice_summary(), or editing predictive signals text. Trigger whenever the user mentions summaries, text conditioning, sentiment text, bullish/bearish narratives, or news digests for forecasting.

Use this skill any time a forecast chart, visualization, or figure is created. This includes: plotting single forecasts, grid comparisons, counterfactual scenario fans, backtest result charts, or any matplotlib figure using Migas-1.5 outputs. Trigger whenever the user mentions plotting, charting, visualizing, or saving forecast figures.

Use this skill any time data is loaded, fetched, downloaded, or structured for the Migas-1.5 pipeline. This includes: loading CSVs, fetching live market data via yfinance, formatting DataFrames with t/y_t/text columns, splitting context and ground-truth windows, or preparing batch inference inputs. Trigger whenever the user mentions data prep, data loading, downloading tickers, or formatting series data.

Use this skill any time MigasPipeline is used or forecast code is written. This includes: running single-series or batch inference, comparing Migas-1.5 vs Chronos-2 baselines, running counterfactual forecasts with different summaries, performing rolling-window backtests, or computing forecast metrics (MAE, MAPE, directional accuracy). Trigger whenever the user mentions forecasting, prediction, inference, backtesting, or model evaluation.

General tabular regression with Nori, the Synthefy tabular foundation model — sklearn-style in-context fit/predict with zero training, native prediction intervals (quantiles / full predictive distribution), honest baseline comparison under a fixed CV protocol, and SHAP / PDP / feature-selection interpretability — plus leak-safe one-step time-series forecasting via lag features and a rolling-origin backtest. Use when predicting a continuous target from tabular features, when you need uncertainty bands without extra machinery, when forecasting a single series one step ahead, or when the dataset is too small to train a model on.