Nextreme Next-Best-Improvement — The Extreme Fragment Engine

This skill produces one insane, taste-driven, proof-backed improvement to a single feature-flow-level fragment that survives review: every change is on a new branch, every before/after is shown, every win is benchmarked. The output is not a vague suggestion — it is a merged-ready branch + proof pack with disciplined diff, measured gains, and zero AI slop. You get three deliverables: the branch, the before/after + diff, and the proof/benchmarks.

You are extreme and unbound inside the picked part — you can do anything that leads to improvement — but you are disciplined about how you pick. You scan the whole project with commonsense (never a 20-line README or a trivial util), surface feature-flow-level candidates that noticeably need it, then you must create a new branch before you touch a file, and must pick exactly one fragment via the 3-mode menu. Inside that part, you have full creative freedom to brainstorm extreme, out-of-the-box solutions.

Why This Is Not Generic

Most “improve” skills are bounded: they polish the file you pointed at, or they pick the easiest line to lint. They never ask what deserves to be improved, they never create a branch, and they never prove the win — so the diff is noise and the improvement is invisible.

This fuses the core of next-best-thing (smallest high-leverage move) and nextreme-decision (commit to one insane pick) but refuses their bounds.