Choose training or context
The context graph is the connected, inspectable context Sphinx supplies to agents. Context mode edits that graph directly; training mode uses your sources in the post-training process to learn context that helps with new tasks.
Write a training prompt
Describe the sources and the goal
Name the material, the intended audience, and the work your agents should become better at. Include realistic examples of future tasks.Identify authoritative sources
Explain which sources govern current practice and which are historical examples. Include effective dates and versions when relevant.Specify the details that matter
Call out definitions, calculation rules, exclusions, policy thresholds, and terminology that agents need to apply correctly. Explain what to leave out.Sphinx learns methods and meaning, rather than memorizing numerical values. Documents containing numbers can teach Sphinx how a calculation works, what a metric means, and where to find the relevant data.A useful rule of thumb: a human expert would not remember the value in cell F88 of a spreadsheet, but would know which spreadsheet to open, what its figures represent, and how to use them. Sphinx learns that expertise. Your agent retrieves the actual values through its existing data or file tools when answering a question.
Write a prompt for a direct context edit
Choose Use files as context when you already know the change you want and the files provide evidence or instructions for that change.- Name the affected concepts or page titles.
- State the correct rule and its scope.
- Explain which source supports the change.
- Keep each request focused on one clear outcome.
What to avoid
- Vague goals. “Learn everything” gives less direction than naming the work agents should improve at.
- Instructions for every processing step. Describe the outcome; let Sphinx carry out the training or edit.
- Unsupported assumptions about your sources. Only describe content and structure you know are present.
- Pasting source documents into the prompt. Attach the files and use the prompt for instructions.
- Unrelated changes in one request. Separate broad training goals from precise corrections when they need different treatment.