| name | llm-model-docs |
| description | Look up current official documentation AND current model lineups for OpenAI, Google Gemini, and xAI (Grok). Use BEFORE adding/changing model IDs, before answering model-capability questions (e.g. "does gpt-4o-transcribe accept a system prompt?"), AND proactively whenever you touch chat/transcription/TTS code paths (provider files, SettingsConfiguration, TranscriptionModels, ChatModelCommandResolver, SpeechService) — compare our current defaults to what each provider has shipped recently, and surface upgrade suggestions to the user if newer GA models exist. AI moves fast — always check live docs first. |
LLM & Speech Model Documentation
Use this skill before doing any of the following:
- Adding, renaming, or changing model IDs in SettingsConfiguration.swift, TranscriptionModels.swift, AppConstants.swift, or anywhere else a model ID is hardcoded.
- Answering user questions about what a model supports (parameters, modalities, context window, pricing tier, GA vs Preview).
- Choosing between model variants for a new feature.
- Verifying whether an old assumption about a model still holds (e.g. token limits, supported parameters, deprecations).
Also use proactively whenever you are working in any of these areas, even if the user did not ask about models:
In those areas, do the lineup check (next section) before finishing the task and proactively tell the user if anything looks outdated. The user wants to hear about new GA models from this skill — not from X.com or coincidence.
Rule of thumb: training data is stale. If you find yourself about to say "model X supports Y" or "the prompt field is limited to N tokens", fetch the doc first. Cite the URL in your answer.
Lineup check (proactive model-currency review)
When you touch one of the code areas listed above, run this short check before finishing:
- Enumerate what we use today. In SettingsConfiguration.swift, the
SettingsDefaults struct is the single source of truth for default selections (e.g. selectedTranscriptionModel, selectedPromptModel, selectedChatModel, selectedMeetingSummaryModel, selectedImprovementModel, readAloudModel). The full per-provider enum is in SettingsConfiguration.swift (PromptModel, TTSModel) and TranscriptionModels.swift (TranscriptionModel).
- Fetch the current model index for each provider whose model appears in defaults or in the enum (see URLs below).
- Compare: for each role (transcription, dictate-prompt, chat, meeting-summary, smart-improvement, TTS), is the default still the best current GA model? Are any enum cases pointing at deprecated/retired IDs?
- Surface findings to the user. Format: "We default to X for role Y. Provider has shipped newer GA model Z — recommended switch / cosmetic update / no change needed." Don't just silently migrate — let the user decide. The user explicitly said: "ich möchte, dass du ab und zu mal schaust, was für neuere Models es gibt … und dann entsprechend Vorschläge machst, falls wir veraltete Models verwenden".
When deciding between "Flash"/"mini"/"lite" variants and the flagship variant, pick by role intent:
- Transcription default → cheap-fast Flash/Lite tier (it runs on every dictation). Flagship is overkill.
- Dictate Prompt / Chat default → Flash tier (price-performance balance, the role most users interact with).
- Meeting summary / Smart Improvement → Flash or Pro tier depending on context length and reasoning needs (long meetings benefit from Pro; quick summaries don't).
- TTS → whichever variant is documented as TTS-capable in current docs.
For each suggestion, name the exact replacement model ID, link the doc URL where you confirmed it, and call out GA vs Preview clearly.
Workflow
- Open the relevant docs URL(s) below with
WebFetch. Ask a specific question (parameter list, GA status, token limits, supported features).
- If
WebFetch returns 403 or the page is JS-rendered, fall back to WebSearch with the provider name + topic + the current year.
- Quote the exact wording from the docs when it matters (parameter behavior, limits, model lifecycle).
- Cross-check with the provider's status/forum if something contradicts what code does or what the user reports — see "Forums & status" below.
Provider-specific docs
OpenAI (Chat, Transcription, TTS)
The app uses these OpenAI endpoints today: /v1/chat/completions, /v1/audio/transcriptions, /v1/audio/speech.
Gotchas worth verifying live before relying on them:
whisper-1 prompt field is capped at 224 tokens (vocabulary hint only).gpt-4o-transcribe / gpt-4o-mini-transcribe prompt field accepts GPT-4o-style instructions (no separate system/instructions field), no documented 224-token cap.temperature is not supported on gpt-4o-transcribe family; stream is.- No separate
system field on the transcription endpoint — instructions go into prompt.
Google Gemini API
This is the canonical Gemini reference for the repo, covering chat, transcription, and TTS.
Pick stable/GA IDs over dated preview IDs (e.g. prefer gemini-3.5-flash over a dated gemini-3.5-flash-preview-* variant) when both are listed. Endpoints in code: TranscriptionModel.apiEndpoint, TTSModel.apiEndpoint; base is https://generativelanguage.googleapis.com/v1beta/models/{model-id}:generateContent.
The app uses the Gemini API (generativelanguage.googleapis.com), not Vertex AI — when reading Google docs, make sure you're on the ai.google.dev site, not cloud.google.com/vertex-ai. Cloud Text-to-Speech (cloud.google.com/text-to-speech/docs/gemini-tts) documents the same/similar TTS models for Cloud/Vertex; use it only as a cross-check.
xAI (Grok)
The app uses Grok via the OpenAI-compatible chat completions interface — see GrokChatProvider.swift.
Gotchas worth verifying live before relying on them:
- Grok model IDs change quickly (e.g.
grok-4, grok-4-fast, grok-code-fast-1). Confirm the exact ID is still listed on the models page before committing.
- The OpenAI-compatible endpoint may not support every OpenAI parameter — check the API reference if a feature behaves differently.
Forums & status (use when docs disagree with reality)
When a model returns unexpected errors, or you suspect a deprecation/rename, check these BEFORE assuming the bug is in our code:
How to use in this repo
- Before changing a model ID: open the provider's model index, confirm the new ID exists and its GA/Preview status, then update raw values in SettingsConfiguration.swift, TranscriptionModels.swift, or AppConstants.swift. Keep or add a comment pointing at the doc URL you used.
- Before answering "does model X support Y?": WebFetch the per-model or guide page, quote the exact wording, and link the URL in your response.
- Verify with the live API, not just the docs. Docs lag behind retirements and silent renames. Use the test scripts in
scripts/:
- scripts/test-gemini-models.sh
- scripts/test-openai-models.sh
- scripts/test-grok-models.sh
They read keys from
.env at the repo root (mode 600, gitignored). Each script tests three buckets: current (must 200), legacy (Gemini/Grok: must still serve via redirect; OpenAI: must 404 to confirm retirement), and candidate (exploratory). Add a model to the right bucket whenever you make an enum change so future runs catch regressions.
- Watch for silent redirects. xAI redirects retired slugs to current ones with HTTP 200 but a different
model field in the response. The Grok test script (scripts/test-grok-models.sh) compares response.model vs requested model — if they differ, it prints "redirected → X". When you see a redirect, the enum case is dead weight: remove it and add a migrateLegacyPromptRawValue mapping.
- After code changes: rebuild and restart via
bash scripts/rebuild-and-restart.sh (per the always-applied rule in .cursor/rules/index.mdc).
- If one provider's script fails uniformly, diagnose credentials first. When every line for that provider shares the same HTTP code (e.g. all Gemini
400), run one probe curl and read error.message before blaming model IDs. Common Gemini cases: API key expired / API_KEY_INVALID (renew in AI Studio, update .env) vs per-model 404 (retired slug — see deprecations). Report OpenAI, Grok, and Gemini pass/fail separately so the user does not think all providers failed.
Anti-patterns
- Don't answer model-capability questions from memory alone — recall is unreliable for fast-moving model details.
- Don't propagate stale assumptions from old code comments without re-checking the doc.
- Don't conflate Gemini API (
ai.google.dev) with Vertex AI (cloud.google.com/vertex-ai) — they document different model surfaces.
- Don't trust unofficial mirrors (third-party API doc sites) over the provider's own pages.
