Less than a day after the tempo run, I started the longest session of the project so far.
The instruction was simple. Keep it easy, ignore pace on the hills and do not add a faster finish. I walked the steeper climbs and tried to keep the effort steady.
The data
Metric
Result
Distance
Approximately 18.8 km
Time
2:03:52
Average pace
6:35/km
Average heart rate
128 bpm
Maximum heart rate
143 bpm
Average power
236 W
Elevation gain
250 m
Temperature
26°C
Perceived effort
7/10
Fuel and fluid
500 ml water, one gel at km 10
The average heart rate was remarkably low for a run of more than two hours on a hilly route. It rose gradually from around 120 bpm early in the run to somewhere near 135 to 140 late in the session, but there was no sharp spike.
What the coach saw
ChatGPT rated the effort management very highly. Walking the climbs kept the session aerobic, and the stable power suggested that I did not suffer a major late collapse.
The contrast between the objective and subjective data was interesting. My heart rate looked very comfortable, but I rated the run 7 out of 10. The previous day’s tempo session probably contributed to that difference.
This is exactly why perceived effort belongs in the log. Heart rate alone might have made the run look almost effortless. It was not.
Fuel practice
I drank 500 ml of water and took a gel at kilometre 10. The gel contained around 21 grams of carbohydrate.
That was enough to finish, but the coach recommended taking more during future runs lasting over two hours. The proposed range was roughly 40 to 50 grams of carbohydrate per hour.
That recommendation sounded high compared with what I was used to, so it would need to be tested gradually rather than introduced all at once.
The week in context
This completed a four run week with intervals, an easy run, a tempo session and a long run. It was the strongest week of the preparation so far.
The coach planned another interval session for Tuesday, but with an important condition. A fifth repetition should only be added if I felt properly recovered.
The plan called for a longer tempo session. I warmed up, ran a sustained faster block and then cooled down. The route was not flat, and the temperature had climbed back to 26°C.
By the end of the tempo section, I felt close to my limit.
The data
Metric
Result
Distance
Approximately 11.0 km
Time
1:04:58
Average pace
5:54/km
Average heart rate
145 bpm
Maximum heart rate
169 bpm
Average power
257 W
Elevation gain
64 m
Temperature
26°C
Perceived effort
7/10
The tempo block covered roughly seven kilometres. Pace varied between about 5:05 and 5:40 per kilometre, partly because of the hills. Heart rate rose gradually through the block and reached the high 160s near the end.
Was it really the limit?
ChatGPT challenged my description of the run. I had said that I reached my performance limit, but I also rated the session 7 out of 10.
Those two statements do not quite match. A true maximum effort would normally be closer to 9 or 10. The data looked more like a hard threshold session that remained under control.
I think that is fair. The final part felt very hard, but I was still able to cool down properly. I had not emptied the tank completely.
What the coach saw
The coach liked the duration of the faster block and the fact that there was no sudden collapse in pace or power. It also noted that the hills made pace a poor guide. A flatter route, or a stronger focus on power and effort, would make future tempo runs easier to control.
The target half marathon pace for a finish around 1:54 is close to 5:24/km. I had already run around or faster than that pace for substantial parts of this workout, but not yet for anything close to race distance.
What came next
The following day’s long run was deliberately kept easy. There would be no fast finish. If my legs felt heavy in the morning, the distance could be shortened.
This was one of the tougher decisions in the plan. Running long on tired legs can be useful, but only if the long run does not become a second hard workout.
Not every useful session needs to feel important while it is happening.
This was a straightforward easy run at 22°C. I felt comfortable, stopped briefly around kilometre two and finished without any drama.
The numbers made it more interesting.
The data
Metric
Result
Distance
8.2 km
Time
51:40
Average pace
6:20/km
Average heart rate
133 bpm
Maximum heart rate
148 bpm
Average power
241 W
Elevation gain
Approximately 50 m
Temperature
22°C
How it felt
Easy
Compared with an earlier easy run, I was around 18 seconds per kilometre faster while average heart rate was about 10 beats lower.
Comparisons between two individual runs are never perfect. Weather, route and fatigue can all change the result. Still, this was the kind of pattern I wanted to see.
What the coach saw
ChatGPT described the run as evidence of better running economy. I was moving faster with a lower cardiovascular cost, and the effort felt easy rather than forced.
The average power of 241 W supported that interpretation. This was not simply a case of slowing down enough to lower my heart rate.
The coach rated the session extremely highly. I would be more cautious about attaching a score to one easy run, but I agree with the broader conclusion. Progress often appears first in ordinary sessions. The same route becomes less demanding. The pace improves without a conscious push.
A planning mistake
The conversation after this run also exposed a weakness in the coaching process. ChatGPT briefly lost track of the agreed four run structure and suggested a week with only three sessions. When I pointed this out, it corrected the plan.
That is worth documenting. A conversational coach can respond to new information, but it can also become inconsistent when the history gets long.
The weekly structure was then clarified as Tuesday intervals, Thursday easy, Saturday tempo and Sunday long.
The next test would be a harder tempo session followed by a long run the next day. That combination was designed to build durability, but it also increased the need to keep the Sunday effort controlled.
I returned to the same route I had used the previous week. This time the temperature was 21°C instead of 28°C.
The difference was immediate. The run felt better, and the numbers improved with it.
The data
Metric
Previous comparable run
This run
Temperature
28°C
21°C
Average pace
6:38/km
6:12/km
Average heart rate
135 bpm
143 bpm
Average power
235 W
249 W
Maximum heart rate
164 bpm
168 bpm
How it felt
Controlled
Good
I ran 26 seconds per kilometre faster and produced 13 more watts on average. Heart rate rose, as it should have at the higher intensity, but the effort still felt manageable.
What the coach saw
ChatGPT called it my best faster session of the previous few weeks. Not because the pace was spectacular, but because pace, power, heart rate and subjective feeling finally made sense together.
Some of the improvement clearly came from the cooler weather. The coach estimated that temperature alone could explain part of the difference. The rest was taken as a possible sign of improved fitness.
That distinction matters. It would be tempting to claim a major breakthrough after one faster run. The more honest conclusion is that better conditions allowed me to show more of the fitness that was already there.
One criticism
My recoveries between the harder sections sometimes involved walking rather than slow jogging. ChatGPT suggested using gentle recovery jogs in future sessions.
That would make the workout more specific to half marathon running. The aim is not only to run fast repetitions. It is also to regain rhythm while still moving.
What came next
The next session was supposed to be an easy run of around 8 kilometres. The coach wanted to see whether the improvement also appeared at low intensity.
That was a sensible test. Faster intervals are encouraging, but improved running economy on an ordinary easy day would be stronger evidence that the training was working.
This was the first long run of the project and the most important session of the opening week.
The plan asked for 17 kilometres at a controlled heart rate. Pace was secondary. The route ran through wooded and hilly terrain, and I decided to walk the steeper climbs rather than force myself to run every metre.
The data
Metric
Result
Distance
17.0 km
Average pace
6:36/km
Average heart rate
133 bpm
Maximum heart rate
156 bpm
Average power
236 W
Elevation gain
209 m
Temperature
25°C
Fuel and fluid
Two dates from km 10, 500 ml water
How it felt
Good
The average pace looks similar to an ordinary easy run, but the route included more than 200 metres of elevation and the session lasted close to two hours. Walking the steepest sections helped keep the effort under control.
Most importantly, I still felt good late in the run.
What the coach saw
ChatGPT focused on the average heart rate of 133 bpm. It remained low for a run of this length, and there was no dramatic rise near the end.
The power output also stayed stable. That suggested I was not fading badly, even though the terrain and temperature added difficulty.
The coach called this the most important session of the preparation so far. Its reasoning was straightforward. Speed over short distances was already visible, but this run provided the first real evidence that my aerobic base could support a stronger half marathon.
It also revised its early race estimate. A finish around 1:54 to 1:56 began to look plausible if the training continued well.
Fuel practice
I drank 500 ml and ate two dates from kilometre 10. That was enough to complete the session comfortably, although the coach suggested testing a little more carbohydrate during future long runs.
This was not just about preventing fatigue. Race day is a poor time to discover that a particular gel or feeding schedule does not agree with my stomach.
My takeaway
Walking the hills did not make this a failed run. It helped turn it into the session it was supposed to be.
That may be the most useful lesson from the first week. Training quality is not measured by how stubbornly I follow a pace. It is measured by whether I create the intended effort and recover well enough to train again.
I ran again the following day. It was still warm, but the conditions felt more manageable than they had during the humid easy run.
This was a quality session with several harder sections. The exact structure is no longer fully recoverable from the screenshots available to this project, so I am keeping the description deliberately broad. One of the rules of this blog is not to fill gaps with invented precision.
The data
Metric
Result
Duration
Approximately 57 minutes
Average pace
6:38/km
Average heart rate
135 bpm
Maximum heart rate
164 bpm
Average power
235 W
Conditions
Warm, but better than the previous day
Compared with the easy run, average power rose slightly while average heart rate was a little lower. The faster sections pushed my heart rate into the mid 150s and low 160s, followed by clear drops during recovery.
What the coach saw
ChatGPT interpreted the pattern as a well controlled session. The cardiovascular response looked good, and the recoveries were clear enough to show that I was not simply accumulating fatigue without recovering between efforts.
The more interesting conclusion concerned my legs. The coach believed muscular endurance was more likely to limit me than cardiovascular fitness. In simple terms, my heart and lungs appeared capable of more work than my legs could yet sustain late in a run.
That diagnosis fitted the larger goal. A faster 5K would be useful, but beating my half marathon personal best would depend more on maintaining form and pace after 15 kilometres.
What came next
The long run was moved to Saturday because I could not train on Sunday. ChatGPT recommended 17 kilometres, controlled by heart rate rather than pace. It also asked me to test taking carbohydrates and water during the run.
The decision to extend the long run after two consecutive running days was not especially conservative. Looking back, it is one of the early examples where the plan could reasonably be challenged.
I decided to follow it. The next session would show whether that confidence was justified.
Two days after the first interval session, the plan asked for an easy run. The pace did not matter. I was supposed to keep the effort controlled and let my heart rate decide how fast I ran.
The weather had other ideas.
It was 28°C and humid. The first few kilometres felt fine, but the run became increasingly difficult as it went on.
The data
Metric
Result
Distance
Approximately 8 km
Average pace
6:42/km
Average heart rate
138 bpm
Maximum heart rate
152 bpm
Average power
230 W
Elevation gain
39 m
Temperature
28°C, humid
How it felt
Very demanding by the end
The split pattern told the story. I started between roughly 6:13 and 6:29 per kilometre. Later kilometres slowed to around 6:50 and beyond, with one kilometre taking more than seven minutes.
The interesting part was that my heart rate did not rise dramatically as the pace dropped. It stayed close to 140 bpm for much of the second half.
What the coach saw
ChatGPT viewed the slower second half as sensible effort management rather than a collapse. I had reduced speed while keeping the internal load fairly stable.
The estimate was that the conditions had cost around 20 to 30 seconds per kilometre. That figure should not be treated as a precise calculation, but the broader point was useful. This was not a good day for judging race fitness from pace alone.
The coach rated the heart rate control positively and did not change the long term goal. It did, however, reduce the intensity planned for the following quality session so that I would still have enough energy for the weekend long run.
My takeaway
This run is a good example of why I wanted a more flexible process. If I had looked only at pace, I might have called it a poor session. The heart rate suggested something different. I had adjusted to the heat without turning an easy day into a race.
It still did not feel easy. That matters too. The data explained the run, but it did not cancel out my own experience of it.
The first session of this experiment was supposed to show ChatGPT what my faster running currently looked like. It also happened to fall on a day when the temperature reached 30°C.
That made it a useful first test. A fixed plan would still have asked for the prescribed pace. A coach looking at the conditions should be more interested in effort, heart rate and whether the repetitions remained controlled.
The session
The original plan called for a warm up, six 800 metre repetitions and a cool down. The surviving record of this workout is incomplete, so I will not pretend that I still have every split. What I can verify is the heart rate, the elevation and the power range during the faster sections.
Metric
Result
Average heart rate
137 bpm
Maximum heart rate
168 bpm
Power during the intervals
Mostly 280 to 300 W
Elevation gain
5 m
Temperature
30°C
How it felt
Hard
The route was almost completely flat, which made the repetitions reasonably comparable. My heart rate rose with each faster section and dropped again during the recoveries. It also crept higher as the session continued.
What the coach saw
ChatGPT interpreted the gradual rise in heart rate as a normal response to the heat rather than evidence that the session had gone wrong. The power remained fairly stable, while pace was more affected by the temperature.
The main conclusion was that I had enough speed for the immediate goal. My ability to hold that speed over longer distances was still the bigger question.
The coach also changed the order of importance for future hot weather sessions. Heart rate and perceived effort would come first, running power second and pace third.
That sounds obvious, but it is easy to do the opposite. A target pace is simple and satisfying. Slowing down because of the weather can feel like missing the workout, even when it is the sensible choice.
What came next
The recommendation was to take a recovery day and run an easy 8 to 9 kilometres on Thursday. The goal was not to hit a particular pace. It was to stay mostly between 125 and 140 bpm and see what pace came naturally.
This first session did not answer whether ChatGPT could coach me to a personal best. It did establish one useful principle: the number written in the plan would not automatically be more important than the conditions on the day.
On October 4, 2026, I plan to run the Cologne Half Marathon and beat my personal best of 1:55. This blog documents an experiment: can ChatGPT help me train more intelligently when it has access to the data and context from every completed session?
In the past, I followed standard training plans. They gave me structure, but they could not react to a bad night of sleep, a heatwave, an unexpectedly hard easy run or a session that went better than expected. This time, the plan is not fixed. After each run, I discuss the result with ChatGPT and use that conversation to decide what comes next.
The experiment
I am treating ChatGPT as a data-informed running coach. After every session, I share the available numbers, explain how the run felt and ask for an interpretation. The next recommendation is based on the accumulating record rather than a generic week from a standard plan.
I will try to follow the advice as closely as is reasonable. If I change or ignore a recommendation, I will document that as well. The interesting part is not a perfect success story. It is whether the process can make sensible decisions when training, recovery and everyday life do not line up neatly.
How the coaching process works
I record heart rate with a Polar upper-arm sensor and use an Apple Watch during the run. The activity is then transferred to Strava. From Strava, I share selected screenshots and values with ChatGPT.
The information can include distance, duration, average pace, kilometre splits, average and maximum heart rate, running power, elevation gain and the available charts. I also add information that is not visible in the activity file, especially perceived effort on a scale from 1 to 10, the weather, recovery and anything relevant about fuel or hydration.
This additional context was ChatGPT’s suggestion. It already matters because the first weeks of the experiment took place during very hot weather in Germany. A pace that feels easy at 18 degrees can tell a different story at 30 degrees.
My starting point
Metric
Starting value
Age
45
Half marathon personal best
1:55
Recent 5K personal best
26:54
Estimated VO2max
47.8
Typical weekly volume
30 to 35 km
Target race
Cologne Half Marathon
Race date
October 4, 2026
Goal
A new personal best
I also shared the heart-rate zones shown in my training app. They needed to be treated carefully because the stored resting heart rate of 75 bpm was outdated and higher than my current value. For that reason, the first assessment used pace, heart rate and perceived effort together instead of treating one automatically generated zone as the truth.
The first assessment
ChatGPT’s initial view was that basic speed was probably not the main obstacle. The recent 5K time suggested enough speed to improve on 1:55. The bigger question was whether I could hold an efficient pace for the full half marathon without losing form or slowing badly in the final kilometres.
The early priorities were therefore endurance, muscular durability and consistent easy mileage. Weekly volume could increase gradually if recovery remained good. Long runs would eventually reach 20 to 22 kilometres, while easy runs had to remain genuinely easy. The first working race target was 1:54, which is roughly 5:24 per kilometre.
The first draft of the plan
Day
Session
Tuesday
2 km warm-up, 6 x 800 m at 5:00 to 5:10 per km with 400 m jog recoveries, 2 km cool-down
Thursday
8 km easy at approximately 6:10 to 6:30 per km
Saturday
2 km warm-up, 6 km tempo at 5:35 to 5:40 per km, 2 km cool-down
Sunday
15 km easy at approximately 6:10 to 6:30 per km, with an optional faster finish
The wider plan also included mobility, easy cycling or walking and two strength sessions. Looking back at the first draft, it was ambitious. Two faster sessions, followed by a long run the next day, can create more fatigue than the individual workouts suggest. One purpose of the journal is to show whether the coaching process notices and corrects decisions like this.
How I will judge the result
The final race time matters, but it is not the only measure. I also want to know whether I can stay healthy, train consistently and recover well enough to complete the important sessions. I will watch whether pace improves at a comparable heart rate, whether long runs become easier and whether the advice responds sensibly to heat, fatigue and missed training.
Before the race, I will record ChatGPT’s final prediction. After the race, I can compare that prediction and the result with the decisions that led there.
The rules
Every completed running session will be documented.
Weather, perceived effort and relevant recovery information will be included.
I will follow the advice when it appears reasonable and safe.
If I change the plan, I will explain why.
Poor sessions and mistakes will not be removed from the story.
Medical advice and clear warning signs take priority over any coaching suggestion.
ChatGPT is not a certified running coach or a medical professional. I remain responsible for every training decision. This blog describes a personal experiment and is not individual training or medical advice.
Privacy and what I publish
Running data can reveal more than pace and heart rate. Exact routes, regular start locations and timestamps can expose where somebody lives or when they are usually away from home. I will therefore publish selected performance data, not a complete copy of my Strava account. (details on data flow)
I will not publish exact start or finish locations, identifying map details, other athletes or direct links to private activity data. Screenshots will be cropped, posts may be delayed and only health information that is relevant to the experiment will be included.
Why document it?
First, Codex makes it relatively easy to turn scattered workout notes into a structured journal. Second, the decisions may be useful to other runners who are curious about data-informed coaching. Third, this is also a practical test of how an AI agent handles a long project in which context accumulates and the next decision depends on what happened before.
What comes next
Each completed workout will get a shorter post with the session, the data, the interpretation and the next recommendation. The first test took place on July 14 in 30 degree heat. By October, I should have a much better answer to the question that started this project: can ChatGPT help me beat 1:55?