most AI disappoints because there is no system behind it. not a tool problem, a systems problem
i learned that building an assistant for my own job. it ran on context docs, layered skills, and a file structure it could navigate
i came from partnerships and sales. now i build these systems for companies and people
@anshuc@RobertBeall17 at scale, & for high quality work, you're going to need a lot more developer access and a real foundation underneath, otherwise it's extremely hard for things to stay consistent for the general user. they don't have files or tools or a system the agent can come back to
@BetaProfiles why can't Apple make a snappier overall experience? don't get me wrong i enjoy the animations, but the extra bounce each one has is so over the top in about 80% of cases
@nickvasiles i have not had an issue with this in the slightest. create the skills and your data/brain agnostic to a harness, and it usually figures it out
@elshayib_ sometimes i wonder how OAI and Anthropic can outlast other businesses with seemingly similar AI products, much grander revenue streams & a more established setup
@gregisenberg this makes things easier and consistent. developing the skills & the plugin is the real challenge. once you have them dialed in, work compounds immediately
no matter how smart the models get, they will never be mind readers without the proper context and pointed direction, and that comes down to the foundation and the architecture of the folders, and the data that is in them
i see most people do this half-assed or not at all, a long prompt and almost nothing in the folders, and then wonder why they get bad results
i've never been a desk setup guy. a good monitor and the laptop, that has been it forever
the past couple months i have been slowly getting new gadgets, keyboard, mouse, better desk, and it has changed more than i expected. remapping the mouse configurations has been a massive boost for day to day work. my eyes have been opened
the build tracker went live on our dashboard today. every plan i put together reports into it while the sessions run, the scope, the diff, the blast radius, and every decision that got made along the way
before this i was calling up specific project docs, reading them inside my harness, and pointing the agents at them. all of that still exists underneath, the difference is there's a visual now that updates itself, so the agents read it to pick up context anytime & the humans can look company wide to see exactly what is built or in flight
the next iteration will be tying it into our internal CRM, so every project is attached to the client for complete efficiency and organization
@nickvasiles i have yet to think of it like this, but there's a lot of truth here
which would explain everyone arguing about benchmark numbers. similar to PC part reviews, useful right up until you run your own workload and get different results & experience
@arthurkatcher i've had a similar experience. tough to wrangle the Anthropic models even with a hardcoded hook. shifted to Grok for coding work and the efficiency and lack of bloat has been worth it (so far). xAI has more coming, so it's a good way into that ecosystem for now
what i keep seeing in the people getting the most out of AI is that they're building systems that compound. one thing at a time, verify it holds, add onto it once the foundation is solid
doing that for a year straight puts you further ahead than any tool you could pick up today
when i plan builds or projects, i have an artifact page that the loop updates as each phase lands. the single source of truth for the AI to reference throughout the build
my own metrics on it, the ones that i find useful for the task at hand. this keeps me fully in the loop & aligned day to day
two lines i put in almost every prompt now
"review this for conciseness without it affecting the quality of the output"
"make sure this is fully verified and working against the test before moving forward"
one kills the over-building, the other stops it from telling me it's done when it isn't
@johnennis we use n8n every day. it's far better for deterministic workflows, which are more reliable and accurate than complex agents
we only reach for an agent when the situation actually needs one