I also use LLMs similarly. As a professional programmer, LLMs save me a lot of time. They are especially efficient when I don't understand a flow or need to transform values from one format to another. However, I don't currently use them to produce code that goes into production. I believe that in the coming years, LLMs will evolve to analyze complete requirements, architecture, and workflow and produce high-quality code. For now, using LLMs to write production-ready applications in real-time scenarios will take longer.
I've been pleasantly surprised by GitHub's "copilot workspace" feature for creating near production code. It takes a GitHub issue, converts it to a specification, then to a list of proposed edits to a set of files, then it makes the edits. I tried it for the first time a few days ago and was pleasantly surprised at how well it did. I'm going to keep experimenting with it more/pushing it to see how well it works next week.
A small portion of my regular and freelance work is translating things from a database to something an application can use. A perfect example of this is creating a model in MVC architecture from a database table/stored procedure/function. I used to have a note or existing code I would have to copy and paste and then modify each and every property one at a time, to include the data types. Not hard stuff at all, but very tedious and required a certain amount of attention. This would have taken me maybe 5 to 20 minutes in the perfect scenario, minus any typos in datatypes, names of properties, etc.
Now I'll do something like this for a table, grabbing the column names and data types:
SELECT
COLUMN_NAME,
DATA_TYPE,
CHARACTER_MAXIMUM_LENGTH,
NUMERIC_PRECISION,
NUMERIC_SCALE
FROM
INFORMATION_SCHEMA.COLUMNS
WHERE
TABLE_NAME = 'Table Name Goes Here'
ORDER BY
COLUMN_NAME;
Then I'll ask my custom GPT to make me a model from the SQL output for my application. I do a quick spot check on the new class and done - the code is completed without any typos in much less time. This kind of stuff goes into production on a regular basis, and I feel about as guilty as I did in 10th grade using a TI-89 for matrix operations, which is zero.
