Fascinating since I found the recent Claude models untrustworthy for writing and editing SQL. E.g. it'd write conditions correctly, but not add parens around ANDs and ORs (which gemini pro then highlighted as a bug, correctly.)
If you aren't already (1) telling Claude Code which flavor of SQL you want (there are several major dialects and many more minor ones) and (2) giving it access to up-to-date documentation via MCP (e.g. https://github.com/arabold/docs-mcp-server) so it has direct access to canonical docs for authoritative grounding and syntax references, you'll find that you get much better results by doing one or both of those things.
Documentation on features your SQL dialect supports and key requirements for your query are very important for incentivizing it to generate the output you want.
As a recent example, I am working on a Rust app with integrated DuckDB, and asked it to implement a scoring algorithm query (after chatting with it to generate a Markdown file "RFC" describing how the algorithm works.) It started the implementation with an absolute minimal SQL query that pulled all metrics for a given time window.
I questioned this rather than accepting the change, and it said its plan was to implement the more complex aggregation logic in Rust because 1) it's easier to interpret Rust branching logic than SQL statements (true) and 2) because not all SQL dialects include EXP(), STDDEV(), VAR() support which would be necessary to compute the metrics.
The former point actually seems like quite a reasonable bias to me, personally I find it harder to review complex aggregations in SQL than mentally traversing the path of data through a bunch of branches. But if you are familiar with DuckDB you know that 1) it does support these features and 2) the OLAP efficiency of DuckDB makes it a better choice for doing these aggregations in a performant way than iterating through the results in Rust, so the initial generated output is suboptimal.
I informed it of DuckDB's support for these operations and pointed out the performance consideration and it gladly generated the (long and certainly harder to interpret) SQL query, so it is clearly quite capable, just needs some prodding to go in the right direction.
Even when I hand roll certain things, it still nice to have Claude Code take over any other grunt work that might come my way. And there are always yaks to shave, always.
I found claude sonnet 4 really good at writing SQL if you give it a feedback loop with real data. It will research the problem, research the data, and improve queries until it finds a solution. And then it will optimize it, even optimize performance if you ask it to run explain plan or look at pg_stat_statemnts (postgres).
It's outrageously good at performance optimization. There's been multiple really complex queries I've optimized with it that I'd been putting off for a long time. Claude code figured the exact indexes to add within seconds (not ones I would have got easily manually).
This kind of thing is a key point. Tell Claude Code to build the project, run linters, run the tests, and fix the errors. This (in my experience) has a good chance of filtering out mistakes. Claude is fully capable of running all of the tools, reading the output, and iterating. Higher level mistakes will need code written in a way that is testable with tests that can catch them, although you probably want that anyway.
