system-prompt-builder

Multi-layer system prompt assembly combines base prompts, identity memories, knowledge banks, and budget state into a prioritized structure. The system truncates memories by category to fit context budgets and loads context in parallel to minimize latency during real-time Convex generation actions.

Parallel context loading fetches user and conversation data simultaneously rather than sequentially during prompt construction. This approach minimizes latency when assembling complex, multi-layer system prompts by retrieving all required contextual components at once before merging them into the final prompt structure.

Context budget management truncates memories by category to ensure the total prompt fits within model limits. When function calling is enabled, knowledge bank prompts are included, while lower-priority contextual memories are truncated first to preserve critical identity and base prompt information.

Yes, the system supports graceful degradation by continuing prompt assembly when optional components like knowledge banks or project context are unavailable. User custom instructions are applied as the highest-priority final layer, ensuring they override base prompts even if other contextual layers fail to load.

Document mode prompts are designed specifically for editing workflows within the AI backend. This mode alters the standard multi-layer prompt assembly to support document-focused generation tasks, maintaining priority ordering and budget awareness while adapting the prompt structure for editing rather than conversational outputs.

Prompt assembly fails when memory stores exceed the context budget if truncation is not properly enabled. The system prevents this by truncating memories by category according to priority, ensuring base prompts and identity memories are preserved while lower-priority contextual memories are reduced to fit the budget.