Training Feedback Loops in 3 Steps: WHOOP, Pictures, and Reps

Training Feedback Loops in 3 Steps: WHOOP, Pictures, and Reps

AI coaching works best when it turns noisy signals into specific decisions. The tradeoff is simple: more data only helps if it changes the next session.

The strongest signal in the KB is not that wearables are useful; it is that Justin Harris is willing to ignore them when they stop improving decisions. In Harry Sims’s thread, the client logged two daily caffeine hits, reported ~8:30 of sleep on WHOOP, and offered to export the data; the actual coaching response that matters was not “collect more,” but a dose and timing judgment based on the athlete’s pattern. The mechanism is a feedback loop: observe, compare, decide, repeat. My thesis is falsifiable and practical: AI coaching improves performance only when it shortens the gap between signal and action, and it fails when it turns coaching into a dashboard with no decision attached.

That is the core tradeoff for coaches. A feedback loop can be built from photos, scale trends, sleep logs, training notes, and wearable exports. But every added signal has a cost: athlete burden, coach time, and the risk of treating noise as insight. The KB gives a pretty clear pattern. When the signal is strong and tied to an outcome, Justin uses it. When it is weak, he discounts it. That is not anti-technology. It is selective use of technology.

1) The first job is to identify which signal actually moves the plan

In David LaMartina’s prep thread, Justin’s note reads like a classic coaching loop: big weight drop, leaner feel, some constipation and distension, likely travel-related, and a plan to add a serving of Dulcolax daily for the final 8–10 weeks of prep. Whether you agree with the product choice is not the point here; the decision logic is. The signal was not “distention exists,” because that is too generic. The signal was “distention rose alongside travel, reduced water intake, and prep-related gut dryness.” The response was to target the bottleneck, not the whole problem.

That is what good feedback loops do. They separate the symptom from the driver. For coaches using AI, the temptation is to collapse everything into one score: readiness, recovery, stress, compliance. But the athlete experience is usually more granular. A client can be training well while digestion is off. They can be sleeping enough while water intake falls. They can be “fine” in a wearable and still be carrying a performance problem that only shows up in the mirror, on the scale, or in the session log.

The practical lesson: if a signal does not change the next decision, it is probably not a coaching metric. It is just information.

2) Wearables are only useful when they answer a specific question

Harry Sims’s log is a good example of the limitation. Two caffeinated drinks, a self-reported stable sleep schedule, WHOOP data, and no reported apnea. That is a lot of data, but data density is not decision quality. The useful question is not “Do we have sleep data?” It is “Does sleep data explain a training issue or suggest a change we would actually make?”

That matters because AI coaching systems can easily reward over-monitoring. If the platform asks for ten inputs every day, users assume ten inputs are better than three. Often they are not. The coach still has to identify the smallest set of signals that changes behavior. For some athletes that might be morning bodyweight, step count, and top-set performance. For others it might be sleep duration, hunger, and pump quality. The right set is the one that predicts the next adjustment.

In the KB, Justin’s style points the same way. He is not described as making decisions because a wearable graph looked dramatic. He is making decisions from a combination of report, visual feedback, and athlete context. That is the mechanism AI should copy: not automation for its own sake, but structured attention.

3) Training feedback loops need thresholds, not vibes

Alex Goracy’s thread gives the clearest example of a useful threshold. Justin observed that the athlete was down 2.1 lb, looked more depleted, and still had stubborn lower-back fat. The response was not “keep pushing harder forever.” It was a tradeoff statement: get as lean as possible to maximize the rebound, but recognize there is a line where extra depletion just restores fullness lost in the diet rather than creating new growth potential.

That is feedback-loop thinking at its best. The decision is not “more or less” in the abstract. It is “how much further is worth it?” In performance coaching, thresholds matter because they prevent infinite tweaking. Without a threshold, every bad day becomes a reason to change something. With a threshold, the coach can wait for meaningful deviation before intervening.

This is where AI can help if it is designed correctly. It can summarize trends, flag deviations, and keep the athlete honest about what happened last week versus what they feel happened. But the model should not be making the threshold for you. A coach still needs rules like:

• If a metric is stable but performance falls, prioritize session output over the metric.
• If bodyweight drops but depletion rises faster than expected, adjust before fatigue compounds.
• If a wearable says “fine” and the athlete’s actual training looks worse, trust the training.

Those are not universal laws. They are decision rules. And decision rules are what make feedback loops coachable.

4) The hidden cost of AI coaching is false precision

The more layers you add, the easier it is to think you know more than you do. That is the biggest trap in AI fitness coaching. A dashboard can make an athlete seem measurable all the time, but measurable does not mean interpretable. If you over-index on the wrong indicator, you can make the next session worse.

The KB examples repeatedly show the opposite behavior: use the client’s signal, but keep the interpretation conservative. A little less water can back someone up for days. Distention may show up before the athlete thinks they are “that” depleted. Sleep averages can look fine while day-to-day quality still differs. These are all examples of useful but incomplete feedback.

So the job of AI in coaching is not to replace judgment. It is to compress the loop. Good systems should help answer four questions quickly:

1. What changed?
2. Is that change big enough to matter?
3. What is the likely driver?
4. What do we do next?

If a platform cannot help with those four steps, it is probably adding surface area, not value.

5) For coaches, the best feedback loop is boring

This is the unsexy part: the best loops are repetitive. Check the same key signals. Compare them against a known baseline. Make one decision. See what happened. Repeat. That is not glamorous, but it is how you avoid turning every week into a reinvention project.

AI can improve that process in three ways:

• It can collect structured reports without making the athlete write a novel.
• It can summarize trends across weeks instead of forcing memory-based coaching.
• It can highlight mismatches between subjective report and objective trend.

But it cannot remove the central coaching tradeoff: every added measurement costs time and attention, and every decision should be justified by what it changes. If the signal does not change the plan, the loop is too long.

The KB does not support the fantasy that better coaching comes from more data alone. It supports a more demanding idea: better coaching comes from tighter control of the next decision. That is the real promise of AI in fitness coaching. Not omniscience. Not automation. Just faster, clearer feedback loops that make the next training choice better than the last one.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/david_lamartina___members-tlssnsjthkmnhfqcscszce25acz_vhdm_x2_xdlpx_i.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/clients/skip_hill___members-b7s-_d0oqqz4a1tadlvuztb0y_vzfqlr9c29tpakweg.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/alex_goracy___members-m1vvgmmbnhazzipv5wqmwwhgjuyotawsompy4kzf6ri.md
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/harry_sims___members-m-iwooxrve6qux4pdtrouhwbkhmvrxoandj1tgzve_m.json
• raw/_consumed/2026-05-26/troponiniq_kb.md