Using Anthropic Claude 4.6 Extended for data analysis

[JaapvanEkris]

I experimented a bit with Claude 4.6 to analyze rowing data. I used the RowingData csv, and I needed to provide it with a base overview of the fields used before analysis could start. I provided Claude with the following data structure:

Field	Unit	Description
Stroke Number	—	Sequential stroke index within the session
lapIdx	—	Lap counter
TimeStamp (sec)	sec	Absolute timestamp
ElapsedTime (sec)	sec	Time since session start
HRCur (bpm)	bpm	Current heart rate
DistanceMeters	m	Cumulative distance rowed
Cadence (stokes/min)	spm	Stroke rate
Stroke500mPace (sec/500m)	sec	Current pace, expressed as seconds per 500m
Power (watts)	W	Mechanical power output
StrokeDistance (meters)	m	Distance covered in this stroke
DriveTime (ms)	ms	Duration of the drive phase
DriveLength (meters)	m	Handle travel distance during drive
StrokeRecoveryTime (ms)	ms	Duration of the recovery phase
Speed	m/s	Current boat speed
Horizontal (meters)	m	Horizontal displacement (same as distance)
Calories (kCal)	kCal	Cumulative calories burned
DragFactor	—	Damper resistance as measured by the erg
PeakDriveForce (N)	N	Maximum handle force during the drive
AverageDriveForce (N)	N	Mean handle force over the drive
Handle_Force_(N)	N	Time-series of handle force sampled across the drive, stored as a comma-separated string — normalised to drive length
Handle_Velocity_(m/s)	m/s	Same structure as above; handle velocity at each sample point
Handle_Power_(W)	W	Same structure; instantaneous power (force × velocity) at each sample point

From my experience, based on this data and a training schedule, Claude could provide insights about drive times, stroke distances, HR drift, leg fatigue, etc.

[Abasz]

Give it the code too, it will help it understand how things are calculated. just be specific when you prompt it.

A few months ago when i was working on the calibration faq for ESPRM I have done a test to see if it can actually calibrate in agent mode for ESPRM since it is able to run the simulation and parse the output.

It looked at the code, the docs, the config (i also told it to see ORM) of previous rowers profiles. Before it started its real work as part of reviewing what is what it asked me what kind of rower do i have on its own!!!! then in about 30 minutes of running the simulation and parsing the data, chamgimg the settings seeing what is improving and what not, it made it like 80% there. It made stroke to get detected distance power there it even realised that we are looking at a water rower and ESPRM has only air rowers and that that the max drag factor needs to be set much higher (it took it some runs to figure out why there are no distance). at some point it inserted print statements to see where it goes wrong, what branch the logic os taken.

I was really impressed.

What i am saying is that if it sees the full picture it understands better what the metrics are and mean.
Then you can ask it to generate a documentation of the fileds specifically for LLM-s and then start a new agent and use that documentation for this analysis.

[JaapvanEkris]

A few months ago when i was working on the calibration faq for ESPRM I have done a test to see if it can actually calibrate in agent mode for ESPRM since it is able to run the simulation and parse the output.

I've been feeding the rowing engine to Claude in pieces, to check the physics/math. It had some interesting feedback which will take some time to process. For example: we assume that drag is air-based, even when it isn't (= magnetic). Finding a way to identify both types of drag (hybrids, including the RowErg) at the same time is interesting. I already ran some initial tests but I couldn't stabilize the algorithm's results (yet).

I was really impressed.

I am with Claude. Other LLM's are less impressive. For me this was an experiment to see how well Claude could act as a coach. It did impressively well, detecting HR drift, detecting rushing the recovery when I get tired, etc.. Seeing signs across sessions indicating an incomplete recovery. I must say I am actually considering adding an API hook.

Then you can ask it to generate a documentation of the fileds specifically for LLM-s and then start a new agent and use that documentation for this analysis.

I was looking at adding a llms.txt, to help LLM crawlers to find their way and steer them in the right direction.

[Abasz]

I am with Claude. Other LLM's are less impressive.

I fully agree. I have very good results with their opus model specifically.

So this llms.txt sor of was the idea behind the FAQ, which I have to complete at some point but I did not have the time to do it for ORM yet (and also I was waiting for 0.9.7's docs update :D ), though I have started (https://github.com/JaapvanEkris/openrowingmonitor/blob/create-faq-from-discussions/docs/faq.md). The important part is that I started mapping out the discussions as many of the community wisdom is located in that and started to ask Claude to create an indexing and a summary per discussion (difficulty was also that there is a mixture on Lars's and your repo...). Then use that as well as the docs ... but its huge and there are obsolete things as well.

But to be honest context window will be less of an issue as time goes by. We are already talking about sub agents and all that which helps a lot with this as the work is split between multiple AI agents with one coordinating. In the past 6-7 month my productivity on ESPRM increased exponentially with them and actually I have never had issues with the context (though I must say that I use github copilot in vscode chat extension where Claude can manage its own context window very well on its own creating "memories").

[JaapvanEkris]

I'm considering adding a MCP to OpenRowingMonitor. Code seems doable, and it would allow people to query ORM directly from an LLM. This way people could train and ask their AI of choice to analyse the data. In part this also solves the problem with data analysis platforms (see #243) as tools like Claude can easily provide reports and even do cross-analysis.

I am investigating Tredict as a potential analysis platform (they integrate with Garmin already) and perhaps they make more speed in the Rowing Data arena than Intervals.icu. At least their GUI is a huge improvement over intervals.icu