An AI tool that physically cannot invent an exercise.

1. The problem

Ask a general-purpose language model for a warm-up and it will happily give you one. It will also, sooner or later, invent a movement that does not exist, describe it with total confidence, and recommend it to someone with a bad knee.

For a fitness tool that is not a cosmetic flaw. Two failure modes matter:

• Fabricated movements. A plausible-sounding exercise nobody can safely perform, or a real name attached to the wrong description.
• Contraindication breaches. The user says “avoid my left shoulder” and the model, three paragraphs later, forgets.

Prompting alone does not fix either. You can ask a model nicely not to hallucinate; you cannot rely on it.

2. What we built

The fix is to move the safety decision out of the model entirely. The model never writes an exercise — it picks one by ID from a fixed catalogue of 114 movements, and the catalogue it receives has already had the unsafe options removed.

The order matters. Contraindication filtering happens before the prompt is built, not after the response comes back. If a user flags a shoulder, every shoulder-loading movement is stripped from the candidate list, so there is no path by which the model can return one — not because it has been instructed not to, but because it was never shown them. Any ID that comes back outside the catalogue is rejected before it reaches the page.

Verified in live testing against the production model: 16 of 16 returned IDs valid, none rejected, and zero shoulder movements returned when a shoulder was declared.

3. Keeping it free, and keeping it free

The tool has no login and no paywall, which makes runaway inference spend the obvious way to kill it. Three independent ceilings sit in front of the model — per-user daily, global daily, and a hard monthly cost cap. Any one of them can halt generation on its own.

That works because the underlying cost is genuinely small. A routine costs roughly $0.0006–$0.0016 to generate, so a £5/month budget covers on the order of 4,000 of them. Cheap models, tightly constrained, beat expensive models loosely supervised — a pattern we reuse across client work.

flowchart TB
    classDef extern fill:#fdfcfb,stroke:#c4873b,stroke-width:1.5px,color:#0a0f0f
    classDef hub fill:#0a0f0f,stroke:#c4873b,stroke-width:1.5px,color:#fdfcfb
    classDef gate fill:#f3f2f1,stroke:#9a9a96,color:#0a0f0f

    W["Your workout
+ kit + niggles"]:::extern
    C["114-movement
catalogue"]:::extern

    F["Filter out
contraindicated moves"]:::gate
    L["Model picks
IDs only"]:::hub
    V["Reject any ID
not in catalogue"]:::gate
    R["Your routine"]:::extern

    W --> F
    C --> F
    F --> L
    L --> V
    V --> R
      

The unsafe options are removed before the model is asked. It cannot return what it never saw.

Everything runs on the same stack as the rest of our work — Flask and SQLite in a container on our own hardware, no build step, no framework churn. Your workout text isn’t stored.

An honest note on reach: the tool works, but it ranks for very little. “Warm up generator” gets around ten searches a month — nobody is looking for this by name. It was built to prove a pattern for constraining AI output, and at that it did its job.