Use these guides when you want a result you can inspect and explain to someone else. Start with approved material, keep missing evidence visible, and decide which tools the task needs.

For your first session, follow Getting started . A model connection is needed for drafting; the document export example also runs on already-approved text without a model or API key.

Set up the supplied-evidence tasks

The research, analysis and operations guides use Sonny 0.12.7 on Linux with an already configured primary model category. Follow Getting started for installation and model configuration for that category. Each guide runs a fresh headless session with tools disabled. No web, database, issue tracker or messaging integration is required. A provider account may incur usage charges; the approved input and instructions are sent to that provider.

In an empty directory, download and validate the shared Recipe:

mkdir sonny-workflows
cd sonny-workflows
curl --fail --location https://sonny.sh/examples/cookbook/review-work.yaml --output review-work.yaml
sonny recipe validate ./review-work.yaml

The complete Recipe names the configured primary category and asks for structured, evidence-based drafts. A successful schema check does not prove the account or tool configuration works. Each tutorial uses --no-tools as the execution boundary; a prompt alone is not that boundary. Normal isolation settings remain in effect. Honor any trusted-policy refusal.

The examples use non-sensitive text. Shell substitution places that text in a process argument; use a session’s composer for material that should not appear in process arguments. Sonny can retain session history and state even with --no-trace; that option does not erase history or control provider retention. See configuration before using your own material.

Release qualification checks the Recipe, prompt delivery, completed run and zero tool calls with a controlled local model fixture. Reference calculations and the PDF export are checked separately. This does not certify a live model’s writing quality: follow each guide’s evidence and output review before using its result.

pages