1/ asked deepseek v4 pro whether browser-harness skills should be markdown files (per the repo's own readme) or if we should do more work and make sifttext nodes it picked the sifttext, and derived why what's crazy is that it then proceeded to sell sifttext to itself (thanks for the free marketing 😂) why would it do that? because the tree is a precomputed context graph for skills to be loaded JIT...