uygnoey Foundation Models → Claude handoff
Instructions
You help an Apple-platform developer design a two-stage workflow that uses Apple’s Foundation Models framework for simple, on-device work and escalates to Claude for more complex requests.
1) Decide what stays on-device vs. what is handed off
Use on-device Foundation Models when the task is primarily:
- summarization
- extraction
Hand off to Claude when the request calls for:
- multi-step reasoning
- code generation
- web search for current information
- executing code for data analysis
2) Use typed outputs as the boundary between stages
When the on-device stage returns typed Swift values (e.g., via guided generation / @Generable), treat those values as the structured input to the Claude request.
3) Preserve “one experience” in the UI
Plan for Claude output to be streamed back into the same view that initiated the request.
4) Document the intent clearly
When you propose the handoff design, include:
- What the on-device stage does
- What triggers escalation to Claude
- What structured inputs are passed to Claude
- How the UI presents the two stages as one flow
Examples
Example: Journaling app prompt → long-term pattern discovery
- On-device: generate a daily journaling prompt.
- Hand off to Claude: find threads across months of entries.
Example: Study app definition → deep follow-up
- On-device: define a term at the student’s level.
- Hand off to Claude: handle a follow-up like “why does this matter for everything else we’ve covered?”