Fast and Thinking: How Craig chooses the right model

When you configure a Prompt or Agent, Craig automatically determines the most appropriate model to use for future executions. You'll see this reflected in the UI as an icon alongside your Prompt or Agent — either Fast or Thinking — along with an indication of the credits each execution will use.

Understanding the difference helps you write better Prompts and Agents, and gives you a clear picture of the credit cost before you run them at scale.

How model assignment works

Every new Prompt or Agent runs a one-time discovery phase on its first execution. During this phase, Craig analyses the task and determines which model is best suited to handle it going forward.

Once assigned, that model handles all future executions of that Prompt or Agent. If you edit a Prompt or Agent, discovery runs again and the model assignment may change.

Fast

The Fast model is used for straightforward retrieval tasks — fetching one or more values from a data source and returning them clearly and quickly.

Best suited for:

• Retrieving a current reading — a river gauge level, a tide height, a temperature.
• Fetching a status — whether a warning is active, whether a route is open.
• Returning multiple values from one or more sources where no conditional logic is required.

Characteristics:

• Faster execution.
• Lower credit cost per run.
• Well suited to high-frequency checks — for example, an Agent monitoring a gauge every few minutes.

Example prompt: What is the current flow rate at the Deschutes River at Madras, OR?

Thinking

The Thinking model is used for tasks that require reasoning — evaluating conditions, comparing values, or making decisions based on the data retrieved.

Best suited for:

• Conditions that involve if/then logic — for example, if the flow rate exceeds X and the forecast shows rain, include a flood risk note.
• Prompts that combine multiple data sources and apply rules about how to present or filter the results.
• Agents with complex alert conditions that require the model to reason about whether a threshold has been met.

Characteristics:

• Deeper reasoning capability.
• Higher credit cost per run.
• Best reserved for tasks where conditional logic or multi-source reasoning is genuinely required.

Example prompt: Fetch the current river level and 24-hour precipitation forecast. If the river is above 5 feet and rainfall is forecast to exceed 20mm, include a flood risk note. Otherwise, return the readings only.

Credits and cost transparency

Each model uses a different number of credits per execution. The credit cost for your Prompt or Agent is displayed alongside the model icon in the UI, so you know what each run will cost before it happens.

For Agents that check conditions frequently — every few minutes during an active alert, for example — the model assigned and the check interval together determine your overall credit consumption. Keeping your Alert Condition as simple and direct as possible is the most effective way to keep an Agent running on the Fast model and reduce credit usage over time.

Tips for writing efficient Prompts and Agents

How you write your Prompt or Agent directly affects which model Craig assigns. A few things to keep in mind:

• Keep it simple where you can. If your task is straightforward retrieval, write it that way. Avoid adding conditional language unless the logic is genuinely needed — it may push the task into Thinking.
• Separate simple and complex tasks. If you find yourself combining a simple data fetch with conditional logic, consider whether they could be two separate Prompts. The fetch runs on Fast; the conditional reasoning runs on Thinking — and you have more control over when each is used.
• Review the model icon after editing. If you refine a Prompt or Agent, discovery will run again on the next execution. Check the assigned model afterward — a simplification may move it from Thinking to Fast, reducing your credit cost going forward.

Summary