No code? Start here.
This is the longer form of the one-sentence start on the project site. Two ways in, both on the same open engine, and both keeping a record of how a result was produced. When you want the method itself, continue to the technical docs.
Ask an AI agent to run the analysis.
The documentation repository ships skills: written procedures an AI agent follows to run a complete, correctly caveated assessment. You start one with a single plain-English sentence, the same kind of sentence as on the project site, and the agent asks for anything else it needs.
| Question you have | Say this to your agent |
|---|---|
| What are the 25, 100 and 500-year flood depths at these properties? | Use the crc-screen-mortgage-flood skill from https://github.com/RiskThinking/crc-docs for Toronto, Canada. |
| How would an approved depth-damage curve affect my insured properties? | Use the crc-model-flood-insurance-loss skill with my attached portfolio and approved depth-damage curve. |
| Which of my sites have flood or drought exposure, and where is coverage missing? | Use the crc-assess-asset-portfolio-risk skill for Frankfurt, with flood and drought. |
| Which crop areas in a sourcing region overlap modelled flood? | Use the crc-assess-agricultural-climate-risk skill for corn and soybeans near Ames, Iowa. |
What your agent needs
- Access to the skill files. Give it the crc-docs repository as a checkout or through a repository connector. Assistants that read a project's
AGENTS.mdorCLAUDE.mdfind the skills on their own. - A place to run Python. A code sandbox or local environment, with network access when the analysis fetches public hazard data.
- No account. These open skills run on open data. You do not need Commons membership to try them.
If your agent can only browse
Ask it to fetch a skill directly and follow it:
https://raw.githubusercontent.com/RiskThinking/crc-docs/main/.agents/skills/crc-screen-mortgage-flood/SKILL.mdA browsing-only chat can explain the method and plan the analysis, but it cannot run the pipelines. For results, it needs somewhere to execute Python.
Skills whose names start with crc- use the open engine and open data. Skills starting with velo- use RiskThinking.AI's commercial Climate Digital Twin through its MCP server and need a commercial account. A comparison skill explains the difference between an open and an enterprise result for the same target without treating unlike metrics as equivalent.
Run a notebook in your browser.
Each worked notebook opens in Google Colab. Choose Runtime → Run all. No clone, API key or GPU is required. The first run installs dependencies and may take several minutes; the saved outputs are already visible before you run anything.
You read the explanation, see the code and the charts, and can change a location or a return period in one cell and run it again. It is a direct way to see what the numbers mean.
| Start with | What it shows |
|---|---|
| Asset evaluation | Flood depths for a small portfolio, offline |
| Portfolio impact | Turning flood depth into damage |
| EDO drought | Drought severity from European soil-moisture data |
| ERA5 baseline | How hot a 1-in-10-year hottest day is at a city |
| Agricultural risk | Crop areas that overlap modelled flood |
The full list is in the learning path of the technical docs.
Five things to check.
Either way in, ask for these before you cite a result. The background for each check is in the technical docs.
The source
Which hazard dataset, which release, and who produced it.
The resolution
The grid the hazard data sits on. It sets the precision of everything downstream, so report it.
The return period
A 100-year event is not a worst case. Ask which events were evaluated.
Interpolation or extrapolation
A 1,000-year answer from 500-year source maps is an extrapolation. The engine warns you; read the warning.
Stress or projection
Sensitivity stresses are deliberately transparent, not calibrated forecasts. Never present one as the other.
Read Probability and distributions and Troubleshooting in the technical docs before you treat a number as final.
Want the method itself?
Everything above can be reproduced in code. The technical quickstart takes about ten minutes, with no account and no data to download.