Search
Daily All In

Edition: English

Athletics & Cycling

Cyclist tests AI chatbot Claude as personalised training coach

Writer Tim Perkin, reporting for Cyclingnews, found that off-the-shelf AI training plans were only useful once fed years of personal Strava data.

AI-generated image illustrating cycling training data and a dashboard concept
AI-generated illustration

A Cyclingnews writer trialled several artificial intelligence platforms to see whether they could replace a human coach, concluding that the technology only becomes useful once it is given a rider's own training history.

Tim Perkin tested ChatGPT, Claude, Gemini and Microsoft Copilot before settling on Claude, which he said could read and analyse CSV spreadsheet files directly. He paid for a subscription to use his full training archive.

His first attempt used a basic prompt: "Build me a 12-week indoor cycling training plan of between 7 and 10 hours a week. I can ride 5 times a week, but not on Thursday." Claude returned a structured 12-week block moving through base, build, peak and taper phases, including an FTP test in week one and a re-test in the final week.

Perkin said that output was usable but generic, resembling plans already available in training manuals or online, because it was not built from any personal data.

To get a tailored plan, he downloaded his training history from Strava, the activity-tracking platform widely used by cyclists. Some early results existed only on spreadsheets, which he uploaded into Strava via a third-party tool before requesting his full archive.

Strava sent a 1.7 gigabyte file, from which only one document, activities.csv, was needed to give the AI enough data to work with. Cyclingnews noted that Strava has since launched an MCP connector allowing direct links to AI assistants, sharing both historic and future data.

Perkin advised riders attempting the same process to keep only their best season of data if file size becomes an issue, rather than uploading everything. He also flagged that activities.csv contains heart-rate data, describing it as personal health information worth checking against an AI platform's privacy settings before upload.

He noted that a smart trainer or power meter is what generates the power figures within the file, and that leaving a ride unrecorded by either device will weaken the reliability of any FTP estimate or power target the AI produces.

Perkin said a properly built prompt needed four elements: a persona for the AI, such as instructing it to act as an elite WorldTour coach; a specific task, for example analysing uploaded data to produce a 12-week plan exceeding previous fitness levels; relevant context; and further detail beyond that, which Cyclingnews did not specify in full.