LLM slop is a skill issue
Three things I had wrong about working with these models, and the process that fixed them
Claude has been feeling dumb lately huh? Maybe, but the deeper issue is how we are using it. While everyone has been telling you that it can one shot a whole NASA Mars mission, often what you get is "Astronauts deleted, relaunch and try again!" The average user's expectation of what these models should do is not in line with how these models function, and as a result, whether the model is good or not becomes irrelevant - every model has constraints.
I am just a token burner, but I think it's worthwhile to share these lessons learned.
The context window is not a bucket
The marketers would have us believe that the more information we stuff in there, the better the outcome. Instead, think of context as a focus override - it is the lever that overrides the model's training, significantly. Even if all of the relevant information is in the context, all of the other "stuff" is affecting output. Claude.md, agents, rules, skills, MCP servers, memory, conversation history - all of it is context, and we haven't even gotten to your prompt yet. What is needed is the optimal amount of context at the most opportune time.
Your prompt is a compression of an infinite idea
You have infinite detail in your head about what the completed task looks like. Your prompt doesn't. That gap is where the model guesses - and when you outsource the hard thinking, you are asking it to pick one solution from an infinite space of solutions. It will do exactly what it is trained to do: implement the most likely one. If you want something novel and complex, you have to own the thinking fully.
The fix is a process
I deleted everything - skills, agents, rules, Claude.md down to 5 instructions. Then I implemented a protocol with a fresh context at every stage:
Facts - Don't tell the model your task yet. Ask about the topic, facts only, no opinions, no inference. Output to a doc.
Research questions - Give the model your task and the facts. Ask it to generate research questions. Output to a doc.
Answer research questions - Give the model task, facts, and questions. It asks you what you can answer, researches the rest. Output to a doc.
Design - Give the model task, facts, and research. Have it dump out everything it thinks the task is. Ask and answer open-ended questions about end goals together. This should read like a document for humans, not an implementation. If you skip through this, there is no value in you or the process - you actually have to know what you are doing. Output to a doc.
Structure - Give the model task, facts, research, design. Have it dump the technical details: architecture, file locations, data shapes, API endpoints. This will expose you for the professional you are. If you don't understand why something is being proposed, brush up. Output to a doc.
Plan - Give the model all of the above. Have it create a vertical plan - atomic logical phases at the end of which you can stop, test, and correct. Think building a house: you don't do all the woodwork first, then all the metalwork. You build in phases that depend on each other, and you measure along the way. Output to a doc.
Implement.
I recreated this as agents and skills - lightweight, repeatable. The facts agent does a top-down overview, spawns sub-agents on deeper concepts, and kills branches that stop being relevant. It's a branch search that terminates on classification rather than depth. The five agents and twelve skills are public: gitlab.com/vladimir.kisselev1/agents-skills.
The way to think about AI is not "everyone can be a lawyer or a doctor now." The world will belong to people who can think deeply about their domain and communicate that to these tools in a way that brings out the best in them. The job has shifted from implementation to architecture - but if you don't understand the implementation, you won't make a very good architect.
I'm reachable at vladimir.kisselev.dev@gmail.com.