AI agents will default to traditional app architecture unless the docs teach the Bitfield translation directly. Give the agent a concrete job, the public names it may use, and the private files it must not import or parse.
You ask an AI agent to add a feature. It creates a store, imports another package, parses a JSON file, or calls an implementation directly. The code may compile, but it trains the app back into the old architecture.The useful output is code that chooses the Bitfield path before the first edit, not code that compiles after recreating the old app shape.The prompt to paste
Classification protocol
Before writing code, the agent must fill this out:If the agent cannot fill this table out, it should not edit yet.Before accepting the edit, compare it against the concrete translation page: Translate traditional code. That page shows the exact store, service import, direct call, shell branch, package-file read, and private UI state mistakes this checklist is meant to catch.
Bad output and corrected output
The shape most agents reach for
The agent reaches for a shared store, direct import, package file read, or framework coordinator because that is what most public examples show.Bad: shared object import
Read the public value
React adapter example
Bad: direct implementation call
Request the work
Bad: consumer parses package files
Read package output
Bad: every state value becomes global
Keep visual state local
Bad: React context becomes product architecture
Use Runtime Kit names
Bad: mixed read and work path
Split reading from running work
Review checklist
Common failures
Required self-report
At the end of generated work, the agent should report this exact shape:Next
- Share state correctly: Share state between packages
- Ask for work correctly: Ask another package to do work
- Use the larger agent page: Build with AI agents