Skip to main content

​
Tips

​
Record project context in custom knowledge

Use custom knowledge for product vision, user journeys, core features, and constraints. What to capture：
  • Product vision： Value proposition and goals
  • User journeys： End-to-end flows
  • Core features： Major modules and capabilities
  • Design rules： Standards and constraints
Benefits：
  • AI keeps referencing this context
  • Consistent outputs
  • Better accuracy

​
Define roles and boundaries

Define roles (admin / user / investor, etc.) so generated code respects permissions. Include：
  • Role types： Admin, user, guest, …
  • Permissions： What each role can do
  • Data scope： What each role sees
Example：

​
Organize when knowledge grows large

Split knowledge by domain, task, role, or process instead of loading everything every chat. Load only what the current task needs; load more when multiple domains apply. Keep context focused and reduce conflicting rules. Advanced: toolize reusable knowledge or retrieve on demand. Start with clear boundaries and accurate content; tooling is optimization, not a prerequisite.

​
Generate a project knowledge base

From implemented features, ask the AI to draft project knowledge. Example prompt：
Typical output：
  • Project overview
  • Feature list
  • Role definitions
  • Design standards
  • Technical architecture

​
Save what works and reuse it

When you find:
  • Wording that works well
  • Structures the AI understands reliably
  • Steps that rarely fail
Save it：
  • Store for next time
  • Add to project knowledge
  • Turn into prompt templates
  • Keep in custom knowledge
Benefits：
  • Faster iteration
  • Consistency
  • Fewer mistakes
  • Reuse proven patterns
💡 Tip： Periodically curate these into your own best-practice library.

​
FAQ

Often too much knowledge is loaded at once for the task.Start with Organize when knowledge grows large:
  1. Split by topic — avoid one giant entry;
  2. Load only knowledge relevant to the current task;
  3. For long docs, read summary/outline first, then add detail in batches;
  4. At very large scale, wrap frequent lookups in tools for progressive loading instead of dumping everything into context.
After reorganizing, retry generation or edits.