CrossFit veteran's GPT coaching setup works per-session but loses long-term programming

rabbitholebeer · reddit · 2026-09-17

A Reddit user with 12 years of CrossFit experience details his attempt to build a long-term adaptive training coach with GPT.

Goal: combine trail/ultra running with CrossFit-style strength, power, conditioning and injury prevention, with the model managing training cycles, remembering recent work, and adjusting daily based on recovery.

Current setup: a Project with multiple chats, high-thinking GPT, deep-research reports, his own training history, and detailed project instructions.

Problem: early programming is good, but over time the model collapses into a small loop — recycling exercises, failing to develop prior blocks, losing longitudinal progression.

Open architecture questions: maintain a training-state doc updated per session? Put periodization in a file instead of project context? Split programming/research/injury/feedback chats? Periodically have GPT audit prior weeks? The core ask: make the model sustain a coherent months-long program instead of treating each day as a one-off workout generation.

Original post →

More from Apps

Apps channel →