Apple foundation models

Apple Foundation Models

Implement Apple Foundation Models with availability-gated UI, short prompts, typed output, and explicit fallbacks. Default to the on-device model; treat PCC and 27-only APIs as opt-in upgrades, not the baseline.

Related skills:

• apple-foundation-models (this skill) — decision points, hard-won rules, shipping checklist
• core-spotlight-ask — Core Spotlight indexing, CSUserQuery, and Ask grounded in Spotlight / search-then-summarize
• native-swiftui — availability/quota/error UI, ContentUnavailableView, native system presentation
• swiftui-project-structure — where feature code, prompt assets, and supporting packages should live
• swiftui-expert-skill — Observation, concurrency, performance, Instruments

Agent workflow

1. Gate by SDK first — Foundation Models is iOS 26+; PCC, ContextOptions, and DynamicProfile are iOS 27+ Beta.
2. Choose the smallest capable model — start with SystemLanguageModel; only justify PCC when 4K context / no reasoning is the real blocker.
3. Gate product UI on availability and locale — use model.availability, supportsLocale(), and a non-AI fallback path.
4. Choose the interaction shape — one-shot session, reused multi-turn session, guided generation, or tool calling.
5. Prefer typed output — @Generable / respond(to:generating:) before raw-text parsing.
6. Keep prompts/program logic separate — compute branches in Swift, then inject only the relevant branch into the prompt.
7. Budget tokens explicitly — prompts, instructions, schemas, tools, tool output, transcript, and reasoning all count.
8. Version prompts by OS/model generation when output quality matters.

Hard-won rules

Model choice

• Default to SystemLanguageModel for summarization, extraction, rewrite/refinement, classification, and short creative generation.
• Do not rely on the on-device model for exact math, code generation, or heavy logical reasoning.
• For tagging/categorization, prefer SystemLanguageModel(useCase: .contentTagging, ...) over a generic free-form prompt.
• Move to PrivateCloudComputeLanguageModel only for a concrete need: larger context, stronger reasoning, or long/complex multi-turn flows.

Session shape

• Fresh LanguageModelSession for single-turn tasks.
• Reuse a session only when the transcript is intentionally part of the feature.
• A session handles one request at a time; serialize requests or check isResponding.
• prewarm(promptPrefix:) is optional latency polish, never a correctness requirement.

Prompting

• Give the on-device model one concrete task per prompt.
• Use short imperative phrasing: “Summarize…”, “Extract…”, “Classify…”.
• Ask for shorter output in the prompt before using maximumResponseTokens.
• Instructions outrank prompts; only place trusted content in instructions.
• If a prompt has several conditional branches, compute the branch in app code and inject only that case.

Structured output

• Use @Generable instead of asking for JSON and parsing text yourself.
• Keep guides short; clear property names are often enough.
• Property declaration order matters.
• If the model needs to “show its work”, give it a dedicated first reasoning field so reasoning text doesn’t leak into the answer fields.

Tool calling

• Use tools for grounding, current/app-local data, privileged framework access, or side effects.
• Don’t use tools when your app already knows the data and can put it directly in the prompt.
• Keep tool descriptions and argument guides short.
• Treat .required tool-calling mode as dangerous unless you define an exit condition.
• For Spotlight-grounded Ask over indexed app content (SpotlightSearchTool, progressive hits, search-then-summarize fallback), follow core-spotlight-ask instead of inventing a parallel stack here.

PCC

• PCC requires the managed entitlement, network access, and quota-aware UI.
• The app must handle usage limits and degraded fallback behavior; do not present PCC as always available.
• If PCC is unavailable or a network-dependent request fails, fall back to on-device when the feature can degrade gracefully.

Product states, not just errors

Design explicit UX for:

• model unavailable / not ready
• unsupported locale
• context exceeded
• tool failure
• PCC quota reached / approaching limit
• PCC service/network failure

Shipping checklist

• Deployment target is iOS 26+ (and 27+ only for PCC/reasoning APIs)
• Availability UI covers available, not-ready, and ineligible cases
• Unsupported locale path is handled
• Session usage is serialized; no concurrent calls to one session
• Structured outputs use @Generable instead of raw JSON parsing
• Tool calling is only used where grounding/actions are needed
• .required tool-calling has an exit condition
• Prompt/schema/tool budget was checked against context size
• Prompt variants are gated for model/OS differences when behavior matters
• PCC path includes entitlement, quota UX, and on-device fallback

Additional resources

• API availability, model/version matrix, locale handling, PCC notes: references/api-and-availability.md
• Prompting, context budgeting, tool-calling traps, prompt versioning: references/prompting-and-operations.md
• Foundation Models overview: https://developer.apple.com/documentation/foundationmodels
• Foundation Models updates: https://developer.apple.com/documentation/updates/foundationmodels