Andy Rowlinson

The good thing with LLMs is you get fast feedback about whether your docs actually work.

I think it will work if we:
a) build polished models with standardised naming
b) put docs against the columns and data
c) have a good bank of example queries
d) start documenting that knowledge living inside our heads

I found them useful for doing boring stuff like picking thresholds for capping averages. They can iterate through five different versions quickly. I am experimenting quite a bit with text-to-sql. I tried indexed memories for the LLMs to pick up the kind of knowledge your talking about.

Thanks, I guess I need to add some claude instructions to the repo then.

One underated feature that I don't see mentioned in the old blog is zooming in on the box area, rather than showing half of the pitch.

Ha 🤣. I don't have a personal claude account yet. There is a million side projects that I am itching to finish with its power.

I am wondering what claude's default library would be if you didn't prompt it to use mplsoccer.

If you are using claude out of interest, did it use mplsoccer?

It appears Bournemouth like to recruit from my thread 🤣

Soler / Rayan

bsky.app/profile/numb...

Well at the least the way I have it structured is around domains:

loyalty
bookings
inventory
...

The agent also suggested having a common one around joins/ data quality.

Each domain has it's own MDs. I imagine you could structure files so it could progressively disclose table schema in a domain.

I am building text-to-sql at the moment for my own use. I decided to use skills, as then you can progressively disclose the datasets available to match the query.

Ooh I also have been thinking about this. My aim is to have a database of SQL examples agents could pull from as the models understand everything apart from our data structure.

Yay you're back! I have learnt so much off you and fastai 🚀

The first comment is "Since you clearly need someone to hold your hand through the basics of software engineering, here is a list of tasks you need to complete before this 'library' is even remotely fit for human consumption. Try to pay attention this time"

The last comment on the TODO list is "Get to work. And try not to break anything else while you're at it." 🤣

I tested something similar, asked the other agent to add a TODO list and it named it "TODO_FOR_SLOPPY_AGENT.md" 🤣

2025 NFL Rulebook | NFL Football Operations Explore the official 2025 NFL rulebook.

There is more details on the NFL pitch dimensions here:

operations.nfl.com/the-rules/nf...

I actually want to make a vertical pitch (see picture) where all the numbers (50, 40, 30, 20, 10) are aligned.

In theory, each digit (5, 0 etc.) should fit exactly in a box of a certain size. This is tricky in matplotlib as the text is defined in points. It also needs to work for different figsize.

I am also trying to add NFL to the existing library.

Something that I have been struggling to in matplotlib is to align text placed at a 90 degree angle so that if you place text at x = 30, y = 10 and new text at x = 30, y = 20 the text is aligned with a baseline: matplotlib.org/stable/galle...

The library is written in matplotlib (python).

I am not sure if this is hard, but port my library mplsoccer to typescript, maybe using d3:

github.com/andrewRowlin...

It doesn't have many proper tests, but it does have these notebooks that run some vizes:

github.com/andrewRowlin...

I also spent a long time writing examples in the main repo.

Mine turns the description into a vector using sentence transformers, and then the assitant will hopefully be able to look up stuff similar to the current query at some point. If it works 🤞

Nice, I was doing something similar with duckdb earlier.

I then asked to give me pax counts for London but it instead used airport like 'LH%'. Actually London Heathrow is LHR, which is pretty close. I am now wondering should I give it access to an airport reference table and whether it should be a markdown skill or the database version.

I asked it to improve the skill as pax has a specific meaning. It adapted its own skill and then it nailed it.

I then asked it to describe pax counts in 2025. It didn't do so well as pax has a specific meaning in Finnair. And it didn't find the spot that I described that in the skill.

It did pretty good and explained about covid drops - something I didn't explicity ask for. I worked on the skill some more by describing all the columns of interest in the table.

I started working on this today. I found this is the claude docs and changed it for working with Snowflake. I started with working with one table only. I hooked up my custom python connector and tried a basic query to explain annual flight patterns. Something I already put as an example in the skill

Anyone got any good examples of using skills to do text to SQL?

We already have sparsely documented dbt yml files, is it worth turning these into skills?

New year, new resolution. I hand craft a lot of data insights and datasets. Hoping to use agents more to help.

Thinking of writing a skill.md for each database schema (dataset) so an agent knows how to join views and tables, the main filters to use, and the key columns. #databs

Ta, I like reading your stuff... just a betting newbie.