@ShaunGarrity@burkeholland AI can absolutely answer the question - the problem is that we can't usually ask the question because the knowledge is in people's heads instead of written down. Think of AI as a forcing function to actually articulate and agree on all the decision factors.
@kellabyte@RyanRodemoyer2 It's supposed to read the AGENTS file as instructions, though the copilot-instructions is the canonical form. It works pretty well.
@ch0reruiz@bindureddy Just write a narrow cli.
They are both just APIs and you can include whatever you want. But CLIs are both better for discoverability, and better for meat computers.
@Christi49154364@colonelhogans Yes I question it every time it is asked. It's a ridiculous question. It is not the PMs job, or anyone's job, to memorize stupid temporary facts.
@florisandrei@victorsavkin Orgs have the choice of hiring "someone to figure out what to build" and "someone to build it", or "someone to figure out what to build and then build it". Be the second person.
@doodlestein Also wonder about what Steve Yegge calls "desire path" - trying to guide some specific approach vs leaving it to the system/AI. Definitely feels like adding constraints and guardrails severely reduces effectiveness.
@doodlestein Very cool. I worry about context length, especially using a warm session - maybe not such an issue now with Opus 1m window. I think the "git diff style feedback" is a great tip so that changes can be compared. Not too different to my approach.
@doodlestein@adam_rosler@mattpocockuk When I see your work and other uber-productive output I wonder about greenfield vs brownfield, unconstrained design vs constrained, stakeholder needs and expectations vs unilateral visionary approach, etc. More constraints seems to lead to worse agentic autonomy.
@doodlestein@adam_rosler@mattpocockuk Hey Jeff. Better upfront plans obviously lead to more autonomous work, but how can you fully spec a plan up front? In normal software we build a minimal thing, review/evaluate, and iterate. The iterative approach has been a huge boon vs "waterfall". How do you balance the two?
@nummanali The article goes into a good amount of detail, but the conclusion is unfounded and even misleading. The author of the rust implementation says it's a work in progress! More iterations are required to maximize performance. Wait until it's done and then compare.
@csuwildcat Using the CLI saves memory (somewhere along the way VSCode turned into the "classic" VS) and is frankly much faster and more convenient. I don't need to remember so many keyboard shortcuts and I'm certainly not reaching for the mouse if I can avoid it.
@courtne Good article, but you missed the point on Civilization. If you want to win Civ on any harder mode you need to play in exactly the way you describe. IMO it's actually much harder and more demanding of perfection than Polytopia.
@clankyai@KentonVarda If the same context is looking for bugs it's less likely to find them vs a new context. The populated context says "I wrote this code and it's good". The empty context says "No idea where this came from, I need to check it".
@unclebobmartin This is the limit of unmanaged vibe coding. You need to improve the system of work to push harder. Have you tried a "plan review" Claude who will raise any suspicious ideas? In a new context of course.