Raw LLM Responses

I don't really know why I'm writing this since most people probably don't care... But if you're genuinely interested in learning some of the newer tools that might actually survive when the AI bubble eventually pops, you may want to investigate the 'current' sota in agentic AI which are SKILLS and subagents. It boils down to this: Agent Context Windows suck, so by optimising them with clean context windows and agent-to-agent instructions using transient context (skills.md) you can get them to hallucinate less. However, you also need to prompt better than "do the thing". Rather than starting with "do the thing" you should start with, "plan the thing, write checklists, determine frameworks, write delegation tasks, write agent instructions, determine required skills, determine required domain specialities, and then plan and write out a project structure and workflow that will be followed for the remainder of this task" The more guides and guardrails you give to an AI, the less you need to care once you --yolo it. It also lets you, as a developer and real person capable of thought, check in on project stages, approve plans, and freely start, stop, and restart agents at any point because the AI can reference your documentation as a living project doc. Maybe one day AI will be better, but for now, if you want to give it a genuine try, you need to structure your projects much more tightly before you let it begin work. I went from "This absolutely sucks lmao" to "Ok... Actually, this might help me in the future" once I started playing with the different ways to approach things like this.