Guided Skills
Skills are packaged expert workflows: step-by-step procedures for the platform's more involved flows, with the pitfalls written down. Composing a causal dataset, validating a twin before trusting it, writing SQL in the platform's dialect: each is a skill the assistant loads and follows rather than improvising.
You do not need to manage them. The assistant is instructed to check the skills catalogue before acting on a request, and loads the right one on its own (load_skill). But you can also invoke one directly.
Invoking a skill yourself
In clients that support MCP prompts, every skill is a slash command. In Claude Code, type /mcp__rootcause__ and the full list autocompletes; for example:
/mcp__rootcause__build-and-validate-twinThis loads the workflow into the conversation and the assistant follows it from there. In clients without prompt support, just ask ("use the build-and-validate-twin skill") and the assistant loads it as a tool call.
The catalogue
Bringing data in
discover-and-register-source
Finding an external or public data source and bringing it in: discover, register, preview, import
import-from-database-connector
Importing tables, collections, or files from an already-connected database or storage connector
extract-documents
Turning documents in the workspace (PDF, DOCX, PPTX, text) into structured sources: lifted tables, per-document extraction, entity mentions
research-to-table
Turning web pages or a research prompt into a workspace table (on deployments with web research enabled)
external-data-for-causal-analysis
Importing external data so it survives causal analysis: column types, anchor identifiers, ontology fixes after import
install-foundation-ontology
Vetting and proposing a curated Foundation Ontology package and whether it will join to your data; the installation itself completes in the platform UI
Shaping data
compose-causal-view
Turning a causal question ("what drives X?") into a valid, analyzable dataset
concatenate-datasets
Stacking same-shaped sources end to end (monthly exports, split files) into one table
find-join-keys
Establishing whether two sources can actually join, by probing real values instead of trusting column names
refine-ontology
Repairing the ontology: wrong join relationships, duplicate concepts, mis-classified identifiers
query-with-anchor-sql
The default for data questions: Anchor SQL over concepts — grammar, grains, metrics, and the one-shot error repair loop
write-datafusion-sql
Raw SQL against the underlying sources, for what Anchor SQL cannot express: quoting, joins, dates, medians, and the dialect's limits
answer-business-metrics
Hygiene for descriptive numbers: one metric definition, semantics resolved before aggregating, coefficients computed or labelled as assumed
Causal modelling
assess-data-suitability
Judging whether the data is sufficient and clean enough to trust a causal result
build-and-validate-twin
Building, training, and validating a Digital Twin, including choosing the right twin type
configure-twin-discovery
Discovery and training settings, and injecting domain knowledge: forced edges, treatment and outcome roles, lags
validate-twin
Judging a trained twin before trusting it: per-node fit, graph structure, causal chains into the outcome
Simulating
run-simulation
Choosing and configuring the right simulation for the question: intervention, counterfactual, optimisation, forecast, explanation, root cause, anomaly
explore-environments
Exploring a multi-environment twin: which regions or segments behave alike, how one slice differs, where the model is weakest
Reporting and building
write-report
Generating a report from workspace evidence, then revising it
build-chart
Configuring each chart type correctly: sankey, histogram, scatter, stacked and multi-series
build-application
Building an interactive Application: module wiring, data binding, and the create, run, fix loop
The catalogue grows with the platform; the live list in your client (/mcp__rootcause__ autocomplete, or asking the assistant to list its skills) is always current.
Last updated