Prompting effectively
Starfire performs best when the request makes the job clear. You do not need a giant prompt, but the model should understand the goal, the relevant context, the constraints, and what a successful answer looks like.
A reliable prompt structure
Use four parts:
- Objective — what should be accomplished?
- Context — what does the model need to know?
- Constraints — what must be preserved, avoided, or limited?
- Output contract — what should the response contain?
Give evidence, not just conclusions
When debugging, include the actual error, relevant code, runtime version, and what changed recently. When analyzing a document, attach or reference the source rather than asking the model to reconstruct it from memory.
Separate durable instructions from task details
If the same behavioral instructions are repeated in many conversations, put them in a custom agent. Keep one-off file names, dates, bugs, or project goals in the current prompt.
State uncertainty requirements
For research or operational work, tell Starfire how to handle uncertainty.
Ask for verification where it matters
Code can be reviewed conceptually without being executed. External facts can be discussed without being searched. If validation matters, ask for it explicitly and use a workflow that has the required tools.
Examples:
- “Validate the build after making the change.”
- “Search current sources and cite every time-sensitive claim.”
- “Compare this against the attached configuration instead of assuming defaults.”
Avoid conflicting instructions
Long prompts often fail because they contain requirements that cannot all be true at once. Prioritize the important constraints and explicitly state what can change.
Use projects for recurring context
Do not paste the same architecture description into every chat if the material belongs to an ongoing project. Put durable project context in Projects or Knowledge and keep the conversation focused on the current task.
A short prompt with the right evidence is usually stronger than a long prompt full of generic instructions.
