validate-skill-against-real-projectlisted

# Validate a skill against a real project A skill reads as correct long after it stopped being correct. Task names invert, a tool's flag turns out to be a GNU extension, a `grep` returns nothing on a layout the author never had. None of it is visible by reading — the sentences stay plausible. Running the skill's own commands is the only check that separates a procedure that works from one that merely sounds like it does. ## Procedure ### 1. Pick a target that matches the claimed scope Read the skill's `description` for the scope it claims, then choose a project inside it. Record what makes this target specific: stack and versions, module count, OS, shell, and whether the project has ever been built. ```bash ls -d */ | head; git -C . rev-parse --show-toplevel # candidates and root uname -sr; echo "$SHELL" # OS and shell ``` Where several projects qualify, **pick the smallest** — fewest modules, plainest layout. A skill claiming "any project of kind X" claims the whole range, and the small end is where generic skills break, because the author's own project was not there. Say which end you tested; one target cannot prove the range. ### 2. Extract every executable claim Go through the skill and list what can be run or checked: - commands, and every flag in them - task, file and directory names it predicts - assertions about a tool's behaviour ("X rejects Y", "this is a GNU extension") - anything phrased as a table of expected names Pro