Every question the AI Features pages answer
The 11 ai features pages answer 55 named questions between them. Every one is listed below, and every link lands on the paragraph that answers it rather than the top of the page.
AI Workout Plan Generation: How to Build One That Ships
- Can I just ask an LLM to write the workout plan?You can, and the demo will look convincing. The problem is what comes back: exercise names that may not exist in your catalogue, invented…
- What should the model return instead of a finished plan?Identifiers and intent, not content. Have it return exercise IDs drawn from a candidate set you supplied, a rep-scheme category picked from…
- Should the LLM decide sets, reps and weights?Have it propose intent, not numbers. Let it pick a rep-scheme category (strength, hypertrophy, endurance, timed) from a fixed enum, then…
- Where in the pipeline does the safety gate run?Before generation, not after. Screen the intake and any free-text the user typed with a deterministic check first, decide which path the…
- How do I regenerate next week's plan without contradicting last week's?Pass the relevant history as structured facts, not as prose, and make continuity a rule rather than a hope. Your code should decide what…
AI Food Logging: Text and Photo Nutrition Entry That Actually Works
- Should the model return calories and macros directly?No. Have it return a food query string, a quantity, and a unit, then resolve that to a row in your food database and compute the macros…
- How accurate is photo calorie estimation?Not accurate enough to present as a precise number, and no honest figure applies across the board. Published evaluations of photo-based…
- Do I need a vector database to log "two eggs and a slice of rye"?Not on its own. Extraction into quantity, unit and a search string is a small structured-output call. Resolution is then a search problem…
- Why did photo logging get worse after I added client-side image compression?Because you probably added a second or third lossy pass. Provider vision documentation states that image compression artifacts are…
- What should I measure to know whether the feature is working?Log the model's proposed food ID and portion, the version the user finally committed, and which candidate they picked if they swapped. That…
Personalizing Your App With a User's Own Wearable Data
- Should I send raw heart-rate or HRV samples to the LLM?Generally no. Serializing days of samples into a prompt asks the model to do aggregation and arithmetic, which is the part models are least…
- Do I need the user's permission to send their health data to an LLM API?For iOS, Apple's App Store Review Guideline 5.1.2(i) says you must clearly disclose where personal data will be shared with third parties,…
- How do I stop the model inventing numbers about the user?Two layers. In the prompt, state that the model may only reference values present in the supplied payload, must not compute new ones, and…
- What should the app do when the user's wearable data has gaps?Handle it explicitly rather than papering over it. Do not silently impute missing days and label the result the same as measured data.…
- Can the model interpret a user's HRV or recovery trend for them?It can describe it. Be careful about letting it explain it. Describing a below-baseline week and suggesting a lighter training block is…
Grounding an LLM in Your Exercise Database
- Do I need a vector database to ground an LLM in my exercise catalogue?Usually not. Vector search exists to solve semantic retrieval over large unstructured corpora. An exercise catalogue is typically low…
- How do I stop the model inventing exercises that are not in my database?Do not try to make it unlikely, make it impossible to persist. Restrict the model to returning an exercise_id, supply the exact candidate…
- Can a JSON schema enforce sensible sets and reps?No. Structured output guarantees the response is valid JSON matching your schema, but the schema layer does not support numeric constraints…
- Should the exercise IDs go in a schema enum or in the prompt?A schema enum is the most direct way to say 'pick only from this list', and enums are supported. The practical catch is that compiled…
- Does the same approach work for a food database?The grounding contract does, the retriever does not. Food catalogues are large, user-generated and lexically messy, so users type brand and…
Writing the System Prompt for an AI Fitness Coach
- What should actually go in an AI fitness coach system prompt?Six things: the role and scope of the assistant, an explicit list of what it may do, an explicit list of what it may not do, how to use the…
- Where should the cache breakpoint go in a coaching prompt?After the parts that never change and before the parts that change every turn. In practice that means the role and scope definition, the…
- Why is "you are a certified personal trainer" a bad first line?Because it invites the model to behave like something holding a credential — asserting rather than hedging, prescribing rather than…
- How do I tell the model which fields it has been given?Enumerate them explicitly and say what to do when one is missing. A prompt that hands over a blob of user context without naming its fields…
- Does prompt caching change how I order the prompt?Yes. Caching is a prefix match over tools, then system, then messages, and any byte change inside the prefix invalidates everything after…
AI vs Rules-Based Coaching: Where a Language Model Actually Helps
- Should an LLM generate the workout plan, or should I just write the progression logic?Write the progression logic. Load progression, set and rep math, volume caps and deload timing are pure functions of data you already…
- Can I ship an AI coach without a rules engine underneath?You can, but you are then trusting a language model with the parts of the product that have correct answers — progression, volume, load…
- Do I have to rewrite my rules engine to add an AI coach?No, and you should not. The migration is additive. First expose the engine's output as structured data, then ship narration — the model…
- Which AI fitness features are lowest risk to ship first?Features that narrate structured data you already computed. Your engine produces the session, the recap, the weekly volume; the model turns…
- Is a rules-based fitness app less capable than an AI one?For anything with a correct answer, it is more capable, because it is right every time and you can prove it. Personalized and algorithmic…
How to Evaluate an AI Fitness Feature
- How do I write tests for output that is different every time?You assert properties of the output rather than equality against a fixed string, because prose has no equality check and the model will…
- Is there a standard benchmark for AI fitness coaching quality?Not that we could find. Our research located no public benchmark, leaderboard or published methodology specific to evaluating AI fitness…
- Do I need LLM-as-judge, or is plain code enough?Start with plain code, and add a judge only for what code cannot reach. If you generate plans by selecting from a vetted catalogue, most…
- How many test cases do I need, and how much does a run cost?Provider guidance on eval design recommends prioritising volume of automatically-graded cases over a small number of hand-crafted ones, so…
- Does an app store actually require adversarial testing?Google Play's AI-Generated Content policy tells developers to test across user scenarios and to safeguard against prompts that could…
Guardrails for an LLM That Gives Fitness Advice
- Can I just put the safety rules in the system prompt?Not on their own. A model asked to police itself in the same call where it is trying to be helpful is subject to sycophancy — the…
- What should an LLM fitness coach do if a user mentions chest pain?Our engineering judgement is a hard stop with emergency guidance, no workout generated and no continuation of the coaching thread — the…
- Should I block pregnant users or users with heart conditions from AI-generated workouts?Constrain, do not refuse. Refusing a pregnant user a walking programme is itself a harm, and the populations most likely to trip a risk…
- Does a 'not medical advice' disclaimer protect me legally?No. It is an app-store expectation and good practice — Apple's Guideline 1.4.1 says apps should remind users to check with a doctor in…
- Does Apple have a generative AI guideline I need to follow?No. The App Store Review Guidelines text contains no occurrence of 'generative' or 'machine learning', and chatbots appear only in…
Choosing an LLM for a Fitness App: Pick by Job, Not by Leaderboard
- Which LLM is best for generating workout plans?Use a frontier-tier model for open-ended plan generation, because instruction-following and constraint retention are what fail visibly on…
- Should I use one model or several?Usually several. A frontier model for open-ended generation and multi-turn coaching, and a small fast model for the high-volume mechanical…
- How much does an LLM feature cost per user?We will not give you a per-user figure, because it depends entirely on your token counts. Cost is dominated by how many tokens you send and…
- Can I self-host an open-weight model instead of calling an API?Yes, and it is the right call when user data genuinely must not leave your infrastructure, since that becomes a fact about your topology…
- What should I check before signing with a model vendor?For a health-adjacent app, these routinely outrank benchmark scores: whether the vendor will contract in writing for no retention and no…
What Does an AI Fitness Feature Cost to Run?
- How much does it cost to add an AI coach to a fitness app?Anyone who gives you a single number for this is guessing. The honest answer is a formula: (input tokens + output tokens priced separately)…
- Why are my input tokens so much higher than I expected?Because the API is stateless. Every request re-sends the entire system prompt, your tool and schema definitions, any retrieved catalogue…
- What is the single biggest lever for reducing LLM cost?Sending fewer tokens. Concretely, in rough order of impact: filter your catalogue in the database before enumerating candidates into the…
- Does photo food logging cost more than text food logging?Yes, materially. Images are billed as visual tokens, and the provider documents the count as ceil(width / 28) times ceil(height / 28),…
- Can I cut cost by summarising or truncating conversation history?You can, and it is one of the most effective levers, but it has a safety consequence you must test for. If a user disclosed a knee injury,…
Structured Output for Workout Plans: Constraining What the Model Returns
- Why not have the model write the plan as text and parse it afterwards?Because the format drifts and a wrong parse looks like a right one. Rep and rest notation changes between calls, parenthetical caveats…
- Constrained decoding already guarantees my schema. What is left to check on the server?Everything that is not shape. Three categories in particular. Referential integrity: each returned identifier must exist in your catalogue…
- Should the exercise identifier be a schema enum built from the candidate set?In our judgement no, even though it would make an invented exercise impossible to emit. Constrained decoding compiles a grammar from your…
- What should happen when the response is cut off partway through the JSON?Treat it as a failure even when the fragment happens to parse. Check the response's termination reason before you parse anything, because a…
- How do I change the plan schema without breaking plans I have already stored?Version it and migrate it like any other persisted format, because stored plans outlive the schema that produced them. Stamp a schema…
Back to the AI Features hub, or see every question this site answers.