Skip to main content
Claude changes a model with one tool, propose_model_update. You need the editor role for the user you connected as.
Try this on a copy of your data first. Every accepted change is saved as a new version. Nothing is edited in place, and earlier runs keep pointing at the version that produced them, but they are marked out of date.

​
What can change

  • Inputs and their estimates
  • Options and the decision framing
  • Declared dependencies between estimates (correlations)
  • The influence diagram: add inputs, calculations and modules, draw or remove arrows, set formulas and outcomes
  • Objectives and risk attitude
Every formula is checked before it is accepted.

​
Examples to ask

  • “Change the adoption estimate to a lowest of 50%, most likely 70% and highest 85%, and say why.” This is a three-point (bounded) estimate.
  • “Add an input called support cost, uncertain, between $20k and $60k, and use it in the cost calculation.”
  • “Set the risk tolerance to 50,000 and rank options by certainty equivalent.”
  • “Generate a decision tree from the diagram using the three biggest events.”
Claude includes a short change note with each update, which you can see in the version history.

​
Correlations

Ask “How will the dependencies in this model be simulated?” and Claude uses review_correlations. It shows each pair’s correlation, whether the whole set is consistent, and the nearest consistent values if it is not. To work out a correlation from data, ask it to use calculate_correlation with paired observations or a connected data source. See Correlation.

​
Check your work

After any change ask Claude to validate the model and list what changed since the previous version. Then open the decision in the browser. It is the same record.
Be specific about shapes: say “likely range” (a 1 in 10 chance outside each end) or “hard limits” (nothing beyond them). They mean different things. See Choosing how to describe an estimate.
Next: Run and compare simulations