There’s a moment every self-coached athlete knows well.

You’ve checked your HRV, you have a rough sense of this week’s load, you know sleep wasn’t great — and you still end up making the call by feel. Hard day or easy day? More volume or pull back? The data is there, but the step from data to decision stays manual. And subjective.

A new study in Scientific Reports shows an approach that automates exactly that step. Not with a chatbot that responds to prompts. But with a system that learns from training data — and develops its own decision-making strategies over time.

What the study actually did

Researchers at AnYang Normal University (China) trained a Deep Q-Network — a reinforcement learning system that learns decisions through trial and error. The idea is straightforward: the agent observes a state, chooses an action, gets a rating. Good decision: positive signal. Bad decision: negative. After thousands of cycles, it develops a feel for what works when.

What makes this interesting: the system wasn’t tested on real athletes — that would be methodologically messy and ethically complicated. Instead, the researchers built a digital twin. A data-driven simulation model trained on physiological and performance data from 25 track and field athletes over a full season. The model learns how different training loads affect HRV, sleep, and performance — and simulates the individual response pattern of a specific athlete.

The agent then trains on that model. It can test as many strategies as it wants without waiting for real training cycles. This solves a genuine methodological problem: you can’t run two training strategies in parallel on the same person. Group studies can do this — but then you lose individual relevance. What works for the group average tells you little about how you personally respond to higher frequency or more volume.

What I find interesting about this

The input data is surprisingly accessible: HRV, sleep quality, current training load relative to recent weeks, weekly performance trend. This isn’t a lab setup — it’s what Garmin, Polar, or Whoop already records.

And the reward structure is well thought out. The system isn’t only rewarded for performance gains — it’s simultaneously rewarded for a healthy physiological state. Optimizing at the expense of recovery costs points. That forces the agent to learn strategies that work long-term, not just produce short-term performance spikes.

This is what I’d call the real difference between “actual AI” and rule-based systems. No fixed rules like “if HRV below 50, easy day”. Instead, a system that learns from an individual athlete’s patterns what makes sense for that specific person in that specific state.

What the study doesn’t show yet

I want to be straight about this: the system was validated on a simulation model, not on real athletes in a prospective intervention study. It learned to beat the model — not proven to improve real-world training decisions.

25 athletes, one season, one sport, one Chinese university. That’s enough to develop a concept and show it works in principle. Not enough to call this a production-ready training coach.

On transferability: the principle works anywhere you have comparable input data. But a new sport means new data, new domain adaptation. In strength training, for instance, the evidence base for HRV as a control variable is significantly thinner than in endurance sports.

What this means for practice — right now

Honestly, not much you can act on immediately. No tool to install tomorrow.

But it points in a direction that I find more interesting than anything I’ve seen in AI fitness apps so far: away from generic recommendations, toward a system that learns from your data how you respond to load. That’s the core idea behind the Prompt Paradox — whoever understands their own data makes better decisions. With or without AI.

What you can do today: take your HRV and load data seriously, spot the patterns, and use LLMs to build training plans tailored to your situation. That’s not reinforcement learning — but it’s the same underlying logic.

The full study is in Scientific Reports (DOI: 10.1038/s41598-026-41946-w) — worth a read if you want to go deeper.

When you ask AI tools like ChatGPT, Gemini, or Claude directly for a training plan, you get an answer. Usually one that looks reasonable.

The problem isn’t the answer — it’s the questions that weren’t asked first.

Your current pace. The injury from last year. How many days you can actually train, not theoretically. Whether you’ve ever done a deload. Whether you hate cardio or love it. Whether you already have a plan — and what’s not working about it.

An experienced coach would want to know all of this before writing anything down. A generic AI prompt asks none of it — and fills the gaps with default assumptions it doesn’t tell you about.

That’s the difference between a plan and your plan.

What the TheFitFuturist Training Plan Skill is

The skill is a free, open-source instruction set for Claude. You upload it once — and Claude becomes a training assistant that asks the right questions before generating anything.

Three goal types are covered: running (5k through ultra), strength and hypertrophy, and mixed training for people who do both.

Mixed is the category most tools ignore. But it’s the reality for most athletes: if you run, you need strength work — for injury prevention and performance. If you lift, you need cardio — for recovery and cardiovascular health. Most apps are either running apps or gym apps. This skill handles both — and understands the interference dynamics between them.

The questions don’t come from nowhere

17 years of coaching experience are baked into the assessment structure. Not in the sense that the skill asks everything a trainer could possibly want to know — quite the opposite. The hard part was figuring out which questions actually matter, and which ones just add complexity and failure points to the model.

The skill asks the most relevant ones — in six conversational blocks, not as a form. You answer naturally, Claude extracts what it needs — and only follows up when something is genuinely missing.

More questions don’t mean better plans. The longer the assessment, the more context the model has to process — and the more can go wrong. Six blocks isn’t a compromise, it’s a deliberate decision for reliability.

What this skill is not

A good coach can’t be replaced by this skill — and that’s not the goal. A real trainer asks the right follow-up questions based on your answers. Questions that come from years of experience — ones an AI might never think to ask, precisely because it lacks that experience. That’s the difference between human and model — and it stays that way.

What the skill can do: for everyone without access to good coaching — whether in running, strength training, or both — deliver significantly more than a generic AI prompt.

What happens next

The plan isn’t a generic template. It comes with a phase structure matched to your goal and timeframe — and a complementary program tailored to your injury history and available equipment.

For runners: heart rate zones with actual bpm values, pacing guidance, strides, interval structure. For strength athletes: progressive overload, RPE guidance, set and rep structure matched to the goal — strength or hypertrophy. For mixed athletes: interference management, so that strength and endurance training don’t sabotage each other.

And then it keeps going.

After each training week, you tell Claude how your sessions went. It reads your training log, recognizes patterns — the same pain twice triggers a plan modification, three times triggers a recommendation to see a physio — and adjusts accordingly.

This isn’t a one-shot output. It’s an assistance system that runs alongside your training.

What I need from you

Before release, the skill was tested against 41 personas: runners, powerlifters, beginners, seniors, people with injury history, people who hate cardio. Every version was scored across 10 categories. Current version: v2.3.3.

But 41 test personas are not 41 real humans with real training histories.

I want to know where it breaks. Where the assessment feels unnatural. Where the plan misses the point. Where the adaptation logic over- or under-reacts.

What I’m asking: Install the skill, run through an assessment, use the plan for two weeks — and tell me what happened. Reply to this post, email me, or open an issue on GitHub. Everything helps.

Nutrition, mobility, and recovery protocols are deliberately out of scope — not because they don’t matter, but because a tool that tries to do everything usually does nothing well. Training planning first. The rest will follow.

This is version 2.3.3, not a finished product. It’ll keep improving — with your feedback.

How to install it (2 minutes)

Works with any Claude account — free or paid. No Project needed. Skills are account-wide.

Prerequisite: “Code execution and file creation” enabled under Settings → Capabilities.

1. Download the ZIP: github.com/ChrisSportiveGrwoth/TheFitFuturist → folder claude-skills/tff-trainingsplan-skill/

2. In Claude: Settings → Customize → Skills → upload the ZIP

3. Start a new chat, describe your training goal — Claude begins the assessment automatically

Use it, break it, tell me what happened.

That’s the only way this gets better.

After 17 years as a trainer, I’ve written more training plans than I can count. For beginners, ambitious athletes, and everyone in between. So when AI tools started generating them, I paid close attention.

And I’ll be honest: I’m genuinely optimistic about what’s possible. A well-prompted AI can produce a solid training plan faster than most coaches. The good ones are a different story.

But most people are using these tools completely wrong. And the results show it.

Garbage in, garbage out.

The quality of an AI-generated training plan depends almost entirely on what you put in. Most people type something like “create a training plan for muscle gain, 4 days a week” — and wonder why the result feels generic.

It is generic. Because you gave the AI nothing to work with.

Here’s what actually needs to go in:

• Basic data: age, height, weight, sex

• Training history and current level — “training for 3 years” is worth more than age and weight combined; it determines volume, exercise selection, and progression

• Current performance numbers — not “I’m intermediate” but concrete figures: bench press 80kg × 5, 5k in 25 minutes; without numbers the AI guesses, and you’ll feel it in the plan

• Goals — end goal and milestones along the way

• Injury history and constraints — separate from general limitations; “I had an L5/S1 disc herniation” is a fundamentally different input than “I don’t have much time”

• Training frequency and session length — 4× per week at 45 minutes is a completely different plan than 4× at 90 minutes

• Equipment and training options — full gym, home setup, bands only

One-time effort, permanent context.

You don’t want to re-enter all of this every time. Set it up once — a ChatGPT Custom GPT with your stats and preferences, or a Claude Project with your data as persistent context. From then on, every training request has your full background automatically. Worth the 20 minutes.

Tell it where to look — and how to think.

Two prompts that consistently improve output:

“Use only peer-reviewed research and established coaching methodologies as your basis.”

Or if you have a specific approach you trust:

“Build this plan following the principles of [Eric Helms / Paul Laursen / your preferred methodology].”

The second one is particularly useful if you already know what kind of training works for you. You’re not asking the AI to invent something — you’re asking it to apply a framework you’ve already evaluated.

That’s the input side. But there’s a deeper problem that almost nobody talks about.

The training data problem.

AI models learn from the internet. That means alongside quality sources — peer-reviewed research, textbooks, evidence-based platforms — they’ve also learned from content written by content producers. People who are excellent at SEO and writing. Not necessarily excellent at training.

I’ve watched this happen in this industry for years. Confident, well-ranked content — written by people with no real coaching experience. The AI doesn’t know the difference. It learned from all of it.

Training knowledge and the basics aren’t optional. Not to write the prompt — but to evaluate what comes back. AI doesn’t produce the correct answer. It produces the most statistically probable one. That’s a fundamental difference. A plan can sound completely plausible, follow a logical structure, and still be wrong for you — or just mediocre. If you can’t tell the difference, you’ll never know.

That combination is everything.

Prompt skill plus real training knowledge. This is what I call the Prompt Paradox: the less you know about training, the worse your AI output will be — and the less you’ll be able to tell. The AI sounds confident either way.

The good news: it’s fixable. In the full guide on TheFitFuturist I walk through exactly how to structure your prompts, which platform features are worth using, and where AI genuinely falls short.

→ How to create an AI training plan that actually works

More on this — and the tools to do it yourself — in the next issue.