These are original editorial templates, not official Google prompts or tested Argon outputs. Availability still depends on the release channel.
Customize before running
- Specify success. Describe an observable result rather than “do a great job.”
- Supply the evidence. Include the files, passages or notes the task depends on.
- Define allowed actions. A text instruction does not grant tool access; configure permissions in the actual application.
- Set the output shape. Use a diff, evidence table or action register that a reviewer can inspect.
- Define missing-information behavior. Require the model to leave unknowns explicit.
Check the result independently
For a patch, run the relevant tests. For research, inspect source locations. For extraction, compare every field with the original passage. Record failure and correction effort rather than judging the response only by how convincing it sounds.
The templates use general task-design principles and make no claim about Argon-specific syntax or optimal parameters. When a task involves tools, confirm the actual product supports them and enforce permissions outside the prompt. A request to stay within scope is useful but is not a security boundary.
Follow the matching use-case guide for a complete workflow. Then use published comparisons to choose candidates and pricing scenarios to budget a test.