Using AI assistants with skforecast

Skforecast provides machine-readable context files so that AI coding assistants (ChatGPT, Claude, Gemini, GitHub Copilot, Cursor, and others) can generate accurate, up-to-date code for time series forecasting.

Quick start: provide the context to an LLM

If your AI assistant can access web pages, paste the following URL into the chat:

https://skforecast.org/latest/llms-full.txt

If your assistant cannot browse URLs, open the file and paste or upload its contents instead.

This single file contains the core API reference, forecaster APIs, workflow examples, and best practices. It gives AI assistants enough context to help you with skforecast while reducing hallucinations about deprecated methods or wrong parameter names.

Example prompt:

Using the context from https://skforecast.org/latest/llms-full.txt, create a ForecasterRecursiveMultiSeries with LightGBM, fit it on a DataFrame with 3 series, and run backtesting with conformal prediction intervals.

IDE integration

If you clone or open the skforecast repository in an AI-enabled IDE that supports repository context files, the assistant can load the relevant context from the repo:

IDE / Tool	File loaded	How
VS Code + GitHub Copilot	.github/copilot-instructions.md	Used as repository instructions by Copilot
Claude Code	AGENTS.md	Read from the project root
OpenAI Codex / Aider	AGENTS.md	Read when the tool supports the AGENTS.md convention
Cursor	AGENTS.md	Can be used as project context when supported or configured

These files contain the same core content: project structure, all forecasters, code style, and testing conventions.

Install skforecast context in your agent

The files above are only loaded when the skforecast repository itself is open. To give your coding agent the same knowledge while you work in your own project, install the skforecast workflow skills or connect a documentation server. Pick the option that matches your tool:

Claude CodeCursor, Copilot, Codex, Gemini CLI and othersManual copyContext7 (MCP)Any other LLM

The skills are published as a Claude Code plugin. Run these commands inside Claude Code:

/plugin marketplace add skforecast/skforecast
/plugin install skforecast@skforecast

The 17 skills become available as skforecast:<skill-name> and Claude loads the relevant one on demand, for example when you ask it to forecast a series or to set up backtesting. Update them with /plugin marketplace update skforecast.

Any agent that supports the Agent Skills format can use the skills. The open source skills CLI copies them into the right folder for your agent (requires Node.js):

# Install in the current project (asks which agents to install to)
npx skills add skforecast/skforecast/skills

# Install for a specific agent, available in all your projects
npx skills add skforecast/skforecast/skills --agent cursor --global

# Install only some skills
npx skills add skforecast/skforecast/skills --skill forecasting-single-series --skill prediction-intervals

# Update to the latest published skills
npx skills update

Without Node.js, copy the skills/ folder of the repository into the skills directory of your agent (for example .claude/skills/ for Claude Code; check the documentation of your agent for the exact location):

git clone --depth 1 https://github.com/skforecast/skforecast.git
mkdir -p .claude/skills
cp -r skforecast/skills/* .claude/skills/

Skforecast is indexed in Context7, a documentation server for coding agents that works with any MCP client. Once the Context7 MCP server is installed in your agent, ask for it explicitly:

Create a multi-series forecaster with LightGBM and backtest it. Use context7 with the library /skforecast/skforecast.

The agent retrieves up-to-date documentation snippets together with a short list of rules that prevent the most common mistakes (deprecated class names, missing frequency in the index, intervals expressed as percentiles instead of quantiles).

Paste https://skforecast.org/latest/llms-full.txt into the chat, as described in the quick start. It contains the API reference and the 17 skills in a single file and works with any assistant.

Note

The plugin, the skills CLI and Context7 read the main branch of the repository, so the installed skills always describe the latest released version of skforecast. If you work with an older version, mention it in your prompt.

What's included in the context

The AI context covers:

• Choice of forecaster — when to use ForecasterRecursive vs ForecasterRecursiveMultiSeries vs ForecasterRnn vs ForecasterFoundation, etc.
• All 9 forecaster types — constructors, fit(), and predict() methods including parameters and defaults.
• Model selection — backtesting_forecaster, bayesian_search_forecaster and other hyperparameter optimization methods, TimeSeriesFold, OneStepAheadFold, and their multi-series variants.
• Statistical models — Arima, Sarimax, Ets, Arar wrapped by ForecasterStats.
• Deep learning — ForecasterRnn with create_and_compile_model, LSTM/GRU architectures.
• Foundation models (zero-shot) — FoundationModel + ForecasterFoundation with Chronos-2, TimesFM 2.5/3.0, Moirai-2, TabICL, TabPFN-TS, TFC-T0, Nori, and TS-ICL backends.
• Feature engineering — CalendarFeatures, RollingFeatures, custom features, and exogenous variables (all built into skforecast, no extra dependency).
• Feature selection — RFECV, SelectFromModel for lags, window features, and exogenous variables.
• Drift detection — RangeDriftDetector and PopulationDriftDetector for production monitoring.
• 17 specialized workflow skills — step-by-step guides for common tasks, loaded on-demand by advanced AI agents.

Workflow skills

Skforecast includes 17 modular skills — self-contained guides that AI agents can load on demand when a user asks about a specific topic. Each skill covers a complete workflow with decision trees, code examples, and common pitfalls.

Skill	What it covers
choosing-a-forecaster	Decision guide: "I have X situation → use Y forecaster"
autocorrelation-and-lag-selection	ACF/PACF analysis with skforecast.stats to choose candidate lags
feature-engineering	CalendarFeatures, RollingFeatures, custom features, exogenous variables
forecasting-single-series	End-to-end forecasting with ForecasterRecursive: data prep, fit, predict, backtest, intervals
forecasting-multiple-series	Global model with ForecasterRecursiveMultiSeries: encoding, dict input, multi-level predictions
foundation-forecasting	Zero-shot forecasting with ForecasterFoundation: Chronos-2, TimesFM 2.5/3.0, Moirai-2, TabICL, TabPFN-TS, TFC-T0, Nori, TS-ICL
baseline-forecasting	Simple baselines with ForecasterEquivalentDate (equivalent-date / seasonal-naive / moving-average) to benchmark ML models
metric-selection	Choosing point, probabilistic, and multi-series aggregated metrics
backtesting-configuration	TimeSeriesFold parameters: refit, gap, fold stride, initial train size
hyperparameter-optimization	Grid, random, and Bayesian search with TimeSeriesFold and OneStepAheadFold
feature-selection	RFECV, SelectFromModel: optional trim of lags, window features, and exog
prediction-intervals	Bootstrapping, conformal prediction, quantile regression, interval calibration
statistical-models	ForecasterStats with Arima, Sarimax, Ets, Arar: auto-ARIMA, seasonal config
deep-learning-forecasting	ForecasterRnn with Keras: create_and_compile_model, LSTM/GRU architectures
drift-detection	RangeDriftDetector and PopulationDriftDetector for production monitoring
troubleshooting-common-errors	Frequent mistakes AI assistants make with skforecast and their corrections
complete-api-reference	Full method signatures and availability matrix for all forecasters

These skills are bundled into llms-full.txt. AI agents that support the Agent Skills spec (Claude Code, GitHub Copilot, Cursor, Codex, Gemini CLI, and others) can also load them individually: see Install skforecast context in your agent.

The table follows the same reading order as llms-full.txt: decide → analyse inputs → build → benchmark → evaluate and tune → refine → specialist paths → operate → reference.

Context files overview

File	Audience	Description
llms-full.txt	Any LLM user	Complete context: API + 17 workflow skills
llms.txt	LLMs with web search	Public index with links to all documentation sections
.github/copilot-instructions.md	Contributors (VS Code)	Auto-injected into GitHub Copilot
AGENTS.md	Contributors (Claude Code, Codex, Aider)	Standard agent context file
skills/	Agents supporting Agent Skills	17 workflow skills, installable with the Claude Code plugin or npx skills
.claude-plugin/marketplace.json	Claude Code users	Publishes skills/ as the skforecast plugin
context7.json	Context7 (MCP) users	Controls what Context7 indexes and the rules it gives to agents

The context files are auto-generated from maintained source files (tools/ai/llms-base.txt, llms.txt, tools/ai/ai_context_header.md, and skills/) to ensure they stay in sync with the library. They are regenerated on every release.

Tips for better results

1. Always provide the context URL — Without it, LLMs may hallucinate methods that don't exist or use outdated API names (e.g., ForecasterAutoreg instead of ForecasterRecursive).
2. Be specific about your forecaster — Mention which forecaster you're using. Parameter names and defaults differ across forecasters.
3. Mention the version — Say "skforecast 0.26.0" so the LLM doesn't mix advice from older versions.
4. Validate the output — AI-generated code is a starting point. Use backtesting or an appropriate holdout evaluation to verify model performance.