Solves the problem of subagents not knowing what context they need until they actually start working on a task. Instead of dumping your entire codebase or guessing what's relevant, this runs a 3-cycle loop: dispatch a broad search, evaluate what you got, refine based on what's missing, repeat. Each cycle scores files 0 to 1 and learns your codebase's actual terminology (like discovering your project calls it "throttle" not "rate limit"). Stops when you have 3+ high-relevance files or hit the cycle limit. The pattern is from ECC and targets multi-agent workflows where token limits make "send everything" impossible but "send nothing" leaves agents blind.
npx -y skills add affaan-m/ecc --skill iterative-retrieval --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
juliusbrussee/caveman
mattpocock/skills
obra/superpowers
forrestchang/andrej-karpathy-skills
vercel-labs/skills