Building the Exercise Database

FoxFit needed an exercise library and started with a dataset of 914 exercises. That’s far too many to be helpful to most people, so we narrowed it down to 150.

This post explains the curation process and how we generated muscle-highlighted thumbnails for every exercise.

The Curation

914 exercises sounds comprehensive, but in practice it was bloated:

• Dozens of minor variations (Dumbbell Curl, Dumbbell Bicep Curl, Standing Dumbbell Curl are all the same movement)
• Exercises with no video demonstrations
• Obscure movements nobody actually does in a real gym
• Duplicate entries hiding behind different names

We filtered aggressively with these criteria:

1. Must have a quality video demonstration from a reputable source
2. Must be something people actually do in gyms
3. One entry per distinct movement pattern
4. Equipment requirements clearly specified

The result: 150(ish) curated exercises covering all major muscle groups and movement patterns. Every exercise links to a YouTube video from a trusted source: channels like ATHLEAN-X, Jeff Nippard, and Renaissance Periodization that explain proper form, not just demonstrate it.

Naming Conventions

Inconsistent naming was a bigger problem than we expected, so we standardised everything:

• Equipment first: “Dumbbell Bench Press” not “Bench Press (Dumbbell)”
• Singular form: “Hammer Curl” not “Hammer Curls”
• No brackets or separators: “Rope Tricep Pushdown” not “Triceps Pushdown - Rope Attachment”

40 exercises needed renaming; the AI did the bulk transformation and we reviewed every result to catch anything that didn’t read naturally.

Muscle Thumbnails

We wanted each exercise to show which muscles it targets at a glance. We started with an SVG diagram of the human body with labelled muscle groups.

The process:

1. Map each exercise’s primary and secondary muscles to SVG element labels
2. Generate a modified SVG for each exercise, with primary muscles highlighted in FoxFit’s brand orange and secondary muscles in a lighter shade
3. Export as thumbnails

All SVGs were generated automatically. The AI wrote the generation script; we provided the muscle mappings and colour specifications.

A few muscle groups weren’t in the original SVG. Abductors, brachialis, rhomboids. We added those regions manually, re-ran the generator, and had complete coverage.

The Result

We ended up with a clean, curated database. Every exercise has consistent naming, equipment requirements, primary and secondary muscle mappings, a video demonstration from a reputable source, and a muscle diagram.

This took a few days of focused work. The AI handled the repetitive parts: transforming names, generating SVGs, building the data structure. We handled the judgment calls: which exercises to keep, which videos to link, which muscle mappings were accurate. Without it, this would have been weeks of tedious data entry.