Skip to content

Skills: how the LLM is briefed

When you call ask(), the LLM does not answer from its own memory of skforecast. Every call includes a set of skills: curated documents about a specific forecasting topic (how to choose a forecaster, how to configure backtesting, how prediction intervals work, the exact API of RollingFeatures, and so on). They keep the answers grounded in the current skforecast API instead of an older version the model may have seen during training.

Skills are plain Markdown files bundled with the package under skforecast_ai/skills/. They are synced from the skforecast repository, so they describe the same version of skforecast the generated scripts run on. The returned AskResult records the skills that were sent (answer.skills), so an answer is always traceable to the documents it was built from.

Skills only apply to ask()

refine_plan() and create_cv() use fixed, dedicated skills (feature engineering and lag selection for the first, backtesting configuration for the second). The selection described here is for ask(), which is the free-form entry point.

Where the skills come from

The skills are the ones skforecast publishes in its own repository, copied verbatim by tools/ai/sync_skforecast_assets.py from the skforecast branch that pyproject.toml pins. They are never edited in skforecast-ai, so a skill always describes the skforecast version the generated scripts run on. The upstream Workflow skills table is the canonical list.

Each skill follows the open Agent Skills standard: a directory named after the skill with a SKILL.md file whose YAML front matter declares name and description, a Markdown body with the instructions, and an optional references/ folder with longer material (API signatures, worked examples). Because the format is standard, any agent that supports it (GitHub Copilot, Claude Code and others) can load the same skills directly from the skforecast repository.

How ask() departs from the standard

In the Agent Skills model the agent reads every description and decides on its own which skill to activate, loading the body only when needed. ask() does not leave that choice to the model: the skills are selected deterministically from the context and the question (see below), and each selected skill is sent whole, SKILL.md and references/ included, trimmed as a block to the token budget. The trade-off is reproducibility: the documentation the model saw is a function of the inputs and is recorded in answer.skills.

Available skills

skforecast_ai.ALL_SKILLS lists the valid names, in priority order (the most foundational first). The table below follows the same order and is checked against the package by the test suite, so it cannot drift from the skills actually shipped:

from skforecast_ai import ALL_SKILLS

print(ALL_SKILLS)
Skill Covers
choosing-a-forecaster Decision matrix mapping data characteristics and requirements to a skforecast forecaster class.
autocorrelation-and-lag-selection ACF and PACF analysis with skforecast.stats, ranking lags by partial autocorrelation, feeding them to lags.
feature-engineering Calendar features, holiday distances, rolling statistics with RollingFeatures, differencing, categorical exogenous variables.
forecasting-single-series ForecasterRecursive and ForecasterDirect: data preparation, training, prediction, backtesting, intervals.
forecasting-multiple-series Global models with ForecasterRecursiveMultiSeries and ForecasterDirectMultiVariate: data formats, encoding, per-series transformers.
foundation-forecasting Zero-shot forecasting with pre-trained foundation models (Chronos, TimesFM, Moirai, TabPFN-TS and others) through ForecasterFoundation: accurate forecasts without training, also for short or new (cold-start) series.
baseline-forecasting Seasonal-naive and equivalent-date baselines with ForecasterEquivalentDate, and how to benchmark a model against them.
metric-selection Which metric fits the forecaster type, the prediction output and multi-series aggregation; configuring metric in backtesting and search.
backtesting-configuration Mapping a deployment scenario (retraining frequency, horizon, data budget) to TimeSeriesFold parameters.
hyperparameter-optimization Grid, random and Bayesian (Optuna) search for single and multi-series forecasters, search spaces, cross-validation.
feature-selection Selecting lags, window features and exogenous variables with RFECV and SelectFromModel.
prediction-intervals Bootstrapping, conformal prediction and statistical intervals; residual management and calibration.
statistical-models ARIMA, SARIMAX, ETS and ARAR through ForecasterStats, Auto-ARIMA, backtesting statistical models.
deep-learning-forecasting RNN, LSTM and GRU models with ForecasterRnn and create_and_compile_model.
drift-detection Monitoring in production with RangeDriftDetector and PopulationDriftDetector.
troubleshooting-common-errors Common skforecast errors and the mistakes LLM-generated code tends to make: deprecated imports, wrong names, data formats.
complete-api-reference Full constructor and method signatures of every skforecast forecaster, backtesting and search function, cross-validation class and preprocessing helper.

Automatic selection

When skills is not passed, ask() selects them in three steps.

1. Base skills from the context. If context carries a profile (a ForecastingProfile or any result), its task_type picks the foundational skills. When the context also carries a plan whose forecaster has another task type (a ForecasterStats or ForecasterEquivalentDate chosen over the recommendation), the skills of that task type are added, and a ComparisonResult with a baseline row adds baseline-forecasting. Without a context, the general choosing-a-forecaster skill is used alone.

Task type Base skills
single_series choosing-a-forecaster, forecasting-single-series
multi_series, multivariate choosing-a-forecaster, forecasting-multiple-series
statistical statistical-models
foundation foundation-forecasting
baseline baseline-forecasting
no context choosing-a-forecaster

2. Keyword augmentation from the question. The prompt is scanned for topic keywords and the matching skills are added:

Keywords in the prompt (examples) Skill added
interval, quantile, conformal, bootstrap, uncertainty prediction-intervals
backtest, cross-validation, fold, refit, gap, evaluate backtesting-configuration
hyperparameter, tuning, optuna, grid search, bayesian hyperparameter-optimization
lag, autocorrelation, acf, pacf autocorrelation-and-lag-selection
rolling feature, window feature, calendar, holiday, differentiation feature-engineering
feature selection, rfecv, feature importance feature-selection
metric, mae, mape, rmse, mase, pinball, coverage metric-selection
lstm, gru, rnn, keras, neural deep-learning-forecasting
chronos, timesfm, moirai, tabpfn, foundation, zero-shot, cold start foundation-forecasting
arima, sarimax, ets, arar, statistical statistical-models
drift, monitor, production, distribution shift drift-detection
baseline, naive, benchmark, equivalent date baseline-forecasting
api, signature, kwargs, "parameters of", "default value" complete-api-reference
traceback, debug, exception, fails, TypeError, ValueError, InvalidInputError troubleshooting-common-errors

3. Conflict resolution and budget. Some skills suppress others whose guidance would be misleading next to them: foundation-forecasting removes the lag, feature and interval skills (a foundation model needs none of them), deep-learning-forecasting removes the single and multi-series workflow skills, and statistical-models removes them together with feature engineering and selection. The remaining skills are sorted by the ALL_SKILLS priority and trimmed to a token budget: a fixed ceiling for hosted providers, and for local ollama: models also the space left in the context window after the role prompt, the rendered context and the room reserved for the answer. Trimming drops the last (least foundational) skills first.

The final list is stored in the result:

answer = assistant.ask(
    prompt  = "Which metric should I use to compare the intervals?",
    context = result,
)
answer.skills
# ['choosing-a-forecaster', 'forecasting-single-series',
#  'metric-selection', 'prediction-intervals']

Logging the selection

The selection and any trimming are logged by the skforecast_ai.llm.skills logger at INFO (trimming) and DEBUG (base, augmented and final lists).

Choosing the skills yourself

Pass skills to bypass the automatic selection. The list is sent as is, in the order given and without budget trimming, so keep it short for local models:

answer = assistant.ask(
    prompt = "How do I tune the number of lags and the window sizes together?",
    skills = ["hyperparameter-optimization", "feature-engineering"],
)

A name that is not in ALL_SKILLS is skipped with a warning in the log; it does not raise. Pass include_reference=True to append the full skforecast API reference to the prompt on top of the skills. It is large, so use it for questions about exact signatures and defaults.

The CLI exposes the same option as a comma-separated list:

skforecast-ai ask "How to set up prediction intervals?" \
  --skills "prediction-intervals,hyperparameter-optimization"

See also