jcb wrote: ↑Mon Jun 29, 2026 7:34 pm
Nortaneous wrote:I've been spending the last few months going down the list of utilities I wish I had the time to write and having Claude knock each one out in an hour or two.
Such as what?
Automated performance profilers, tools for extracting information from performance traces, Puppeteer automation, scraping GitHub issue trackers for bugs of certain types to build an input corpus for testing, etc.
Travis B. wrote: ↑Mon Jun 29, 2026 8:42 pm
The only task for which I have ever thought "maybe I wish I could use an AI for this" is software design document generation, because SDD generation is a dreadful, unloved job and no one who actually cares about things really looks at SDD's anyways, so SDD's get very out-of-date because no one bothers to update them, but at the same time it has little impact if an AI hallucinates while updating them because, again, no one really looks at them.
Keeping SDDs in sync is something they'd be good at, although it would take a few iterations of human review to get a good skill written.
Having an AI work on a complex codebase is currently a misapplication. They can
navigate in them, but they don't have the tacit knowledge (and getting tacit knowledge written down is a highly nontrivial project) and they're conservative to the point of refusing to ever refactor. They're good at one-off scripts, "insert tab A into slot B" glue code, and information-dense scutwork (our CTO was opposed to browser automation testing on the grounds that it's unmaintainable, which was defensible until Nov 2025), and fast enough that I'll sometimes throw reviews into them as a quick first pass.
IMO it's likely that AI will become capable of building and maintaining complex greenfield projects before it becomes productive at a senior engineer level in pre-existing codebases, which have both the previously mentioned "tacit knowledge debt" and a totally different pipeline. Generation is cheap and unreliable, so
verification is the bottleneck. It's even possible that, once we know more about maintainable AI-native development workflows, it'll actually be cheaper to port the legacy codebase with extensive differential testing, as with the Bun Rust port. (This was talked about a lot, but it's practically the ideal case for an agentic development megaproject: differential testing -- running the Zig original against the Rust port and asserting that they do the same thing -- provides an ideal oracle, porting to Zig-accented Rust with huge unsafe blocks to clean up later is what a team of human developers would do anyway, and after the Anthropic acquistion the Bun developers would've had access to unlimited Mythos tokens.)
Duaj teibohnggoe kyoe' quaqtoeq lucj lhaj k'yoejdej noeyn tucj.
K'yoejdaq fohm q'ujdoe duaj teibohnggoen dlehq lucj.
Teijp'vq. Teijp'vq. Teijp'vq. Teijp'vq. Teijp'vq. Teijp'vq. Teijp'vq.