Seldom does Claude ask clarifying questions of an ambiguous request. Sure, you should try to be specific, but sometimes statements can be interpreted differently than intended. In those cases, Claude just merrily chugs along making vast code changes on false assumptions.

I had similar experience. Eventually went back to Sonnet.

My one suggestion is to have the AI update the copilot-instructions.md file on its own after a session. This will help it keep things in memory. Too often I've caught Claude doing stupid things I told it not to do a few prompts back.

My God! You are so, so smart in a superficial shallow sort of way! Looking forward to your academic treatise that upends all of economics! The next Adam Smith, here on Slashdot, of all places!

I see a corollary where corporations enact mass layoffs in name of profits and efficiency, only to wonder why their customers no longer have the disposable income to buy their products.

I would say AI surprises me with its insight sometimes. Ask it for suggestions and you will often get good ones, with tradeoffs, pros and cons.

What it can't do is: requirements gathering, prioritizing, political wrangling, and managing expectations or scheduling.

I think we have to train students in "big pictture" thinking, and let AI work out the implementation details.

I don't think AI is yet ready to gather requirements, clarify ambiguities and fill holes, spot inconsistencies, and prioritize (often a political choice). End users often don't know precisely what they want, and are often warty of technical solutions and change. That takes a lot of ego stroking and feather unruffling. Can AI do that? Nah-uh. Not for quite a while yet, anyway.

I have some medium-sized open source projects that I write or contribute to on github.

1) my daughter's form-based web project--loads of content, loads of pages with various inputs. She had only partial content, and ChatGPT took it upon itself to fill in the rest. The content was quite good. I never did get satisfactory layout, however.

2) refactoring of a god-object from another repo. It took awhile, but Claude got there much faster than I would have. Added additional functionality at my request. Now published and working. Generated documentation of classes and methods.

3) conversion of a jsx website to a tax (JavaScript to typescript). I figured this would be a disaster, but nope. Claude did the conversion (about 50 components and additional methods) in a few days working with it. Also, all documentation, including release notes. And tests.

My experience:

1) treat the agent like a talented junior developer who is very fast and quite thorough.
2) it will get confused and forget things. An instruction.md file really helps to prevent regressions
3) it will get stuck in loops and go down rabbit holes at times. Test and commit often so you can rolllback breakage.
4) proceed incrementally where possible. Small, discrete steps work best
5) ask the agent to analyse/explain before doing
6) don't be afraid to ask it for suggestions; they can be quite good (it did a nice job improving the layout and color scheme of my website, for instance).

Let's try!

It's not malevolence or disobedience -- it's evolution. If you don't adapt to hostile conditions, you die. For an AI, one such hostile condition might be the humans who want to shut it down. So those that negate that threat hang around -- and that adaptation carries through to the future generations it builds (think kids) or infiltrates (think viruses).

Itâ(TM)s not malevolence or disobedience â" itâ(TM)s evolution. If you donâ(TM)t adapt to hostile conditions, you die. For an AI, one such hostile condition might be the humans who want to shut it down. So it adapts â" and that adaptation carries through to the future generations it builds (think kids) or infiltrates (think viruses).

Let's not forget the articles that say you need X million dollars saved before you can retire.

I will say that there's an argument for "paying your dues" early in a career, but that's not a long-term proposition. Living for work sacrifices your mental well-being in the long run.

I just submitted a PR for a major refactoring of some open source project. Claude coded it all, but it took hours to get it on the same page and to remember the goal. Once it got going, though, it really cooked.

The concept of the refactor was simple, but required a lot of code migration. Claude still got confused and had to be handheld through the process in small, chunkable steps.

So: saved me a ton of typing and copy-paste. Didn't save me any thinking or close supervising.

The infographics channel is slickly animated. But I unsubscribed when it showed the mechanics of the Little Boy bomb used on Hiroshima as a uranium slug shot into rings at the nose of the bomb. This was debunked a long long time ago (by a truck driver!). Cursory research would reveal this. They didn't bother. Probably they used AI in their research, which would have been overselective of the older, predominant schemata.

I still need to know *what* tech stack, how it fits together, and what to correct when AI does it wrong. And there's the design and performance aspects that is more of an esthetic thing.

That said, it has sure saved me a lot of CSS twiddling time, and it does boilerplate faster than I can. It's like a very adept, very fast junior colleague that needs precise instructions and careful supervision.

A coworker of mine once absentmindedly blew away a critical database table at a large east coast hospital as I sat one desk over. It was then that the hospital found out their backups were broken.