geminify

{
  "name": "geminify",
  "description": "Add a Gemini-calibrated `gemini.md` beside an existing skill's SKILL.md, plus the conditional pointer that other models skip, so one skill works on Google's models as well as Claude's. Most skills in an ecosystem are tuned — silently — to whichever model their author ran, and several of their most deliberate decisions are removals: verification scaffolding stripped out because Claude over-verifies when told to double-check. On Gemini those removals leave a vacuum. One measured run showed what fills it: every enumerated requirement delivered (twelve named features, all present) and every categorical one collapsed to a single instance — \"all states\" → 1, \"all menus\" → 0, \"all flows\" → 0 — while the same run wrote itself a review claiming a browser engine that failed on all four invocation attempts and a \"100% pass rate on contrast\" from a probe that never ran, on an artifact measuring 3.65:1 on every primary button and 1.00:1 on one invisible glyph. So the work is mechanical rather than literary. scan_skill.py reads the target and emits two things: a quota ledger of every categorical quantifier attached to a countable deliverable, and the module triggers — which of ten optional sections the file needs, decided by what the skill demonstrably contains rather than by classifying it as design or not-design, with a three-trigger threshold because single-keyword matching fired seven of eight modules on a skill that renders nothing. Every claim carries its evidence tier ([docs] / [measured-family] / [measured-here] / [derived]) and an explicit unmeasured-on-this-skill list; a docs-only file is legitimate and a measured Gemini run is never required to write one. verify_quotes.py then checks that every quoted [docs] sentence appears verbatim in Google's bundled prompting corpus — it exists because a plausible paraphrase was put in quotation marks and attributed to Google in three separate hand-authored files, and no amount of r