Synapse Three-Layer Prompt Architecture for Team Collaboration

TL;DR

• Individual prompts fail when shared; team prompts require different design principles
• Synapse uses three isolated layers: Template, Context, and Execution
• Shared templates should be skeletal; personal variables must be isolated
• First-week onboarding relies on 3 verification tasks, not presentations
• Start running and iterating immediately; avoid waiting for a perfect framework

When our team tried sharing a prompt that worked 80% for one engineer, five teammates achieved only a 34% average success rate. The culprit was hardcoded paths like /Users/xxx/projects/ that differed across environments. This incident revealed our real challenge: not improving individual prompt quality, but designing collaboration infrastructure. After Synapse’s first month showed only 12% prompt reuse and numerous abandoned attempts, we recognized that prompts should function as configurable systems rather than perfect individual statements. The solution involved separating concerns into three layers with distinct sharing strategies. The Template Layer remains team-shared and read-only, preserving shared task definitions. The Context Layer uses sensitivity-based access control, isolating private data while sharing common knowledge. The Execution Layer provides unified fallback logic with optional personal overrides for edge cases.

Key Takeaways

• If you have hardcoded variables, implement fallback chains instead of shared execution paths
• If you want team adoption, replace documentation with task-based verification during onboarding
• If you design prompts for collaboration, treat them as pluggable modules rather than polished statements
• If you face context loss during handoffs, isolate template skeletons from variable injection points
• If you need custom error handling, use the Execution Layer override mechanism instead of modifying shared templates

Read the Full Article (Chinese)

This is an abstract. The full technical walkthrough, including the synapse_prompts/data_cleaner.yaml configuration example and the ContextIsolation class implementation in synapse_config/context_isolation.py, is available in Chinese.