NICOLAY GEROLD

AMP is a distributed system: an engine, a proxying server, sandboxes, and the local CLI. To debug one problem the agent has to aggregate logs across all of them. The local codebase stores the running logs and the root AGENTS.md tells the agent where they live, so it can reproduce an issue, read the logs, and judge whether the reproduction actually worked.

When a failure keeps recurring, Nico gives the agent a way to see it rather than another paragraph of instructions. Storybooks, debug panes, a focus inspector: "it's often better to actually give it tools or ways where it can actually really debug it and figure out what's going on." That removes a whole class of mistake, guessing the root cause, and leaves the human reviewing design and placement.

When a thread goes wrong, Nico turns the failed conversation into evidence. The skill reads the thread, finds each moment of friction, and traces it to its source: the system prompt, a tool definition, a loaded skill, or an AGENTS.md file. It sorts each one into a failure type (steering, undo, repetition, confusion, wrong tool, missing context, under-agentic, over-agentic) and proposes a concrete fix.

The choice that makes it useful is the bias toward deletion. "I tell it to always default to removing because agents always like to add new stuff instead of removing things." It is written for AMP, so it assumes the read_thread tool and ~/.config/amp/AGENTS.md. Swap in your own tool names and paths and it carries to any harness; read it first, then make it yours.