Step 4 of the commit-crafter workflow is the only instruction in SKILL.md with a machine on the other end of it. The agent shells out to `python scripts/check_message.py --file message.txt`, and step 5 tells it to fix violations and run the validator again [3]. That loop terminates on a process exit status and nothing else. The tutorial's guidance is to keep the script deterministic and return meaningful exit codes [2], which is correct, and it is also where the design work is handed to you: no mapping from failure class to code is given. The visible snippet returns a list of Violation records, and both examples carry the severity "error" [8]. Add a warning severity and you decide whether warnings exit nonzero, and the retry loop will be exactly as reliable as that decision.
Then count what the check encloses. The validator rules on subject format, maximum length, the blank-line separator and the breaking-change footer [6]. The Rules section lists four prohibitions: never stage files yourself, never commit without explicit approval, never add an AI attribution footer unless asked, and read the reference files when the type or format is unclear [5]. Score those against a function whose only input is the message string, and one of the four is checkable, the footer [17]. The other three describe actions and reading habits. They stay prose, as does the separation the tutorial itself calls out between drafting a message and running git commit [10], and the approval gate at step 6 [4].
That is the shape of the problem rather than a flaw in the write-up. A function that receives a string can only rule on the string. For "never commit without approval" to be enforced rather than requested, the return value has to sit where the commit happens: a script that owns the git invocation and hands back a code, or a hook that refuses. Judgment is what belongs in prose because it genuinely varies, and the source leaves the feature-or-fix call to the model, which is the correct half of the split [9].
Cost to adopt the validation path is close to nothing. Standard library only, no package install step [7]. The stated Python 3.12 floor is attached to the repository's CI example rather than to the validator itself [13], so an older interpreter is plausible, though that is my reading of the scope and not something the tutorial promises. The bill arrives upstream: there is no stable release tag and the examples are checked against current main [14], so copying the structure is cheap and tracking it is a maintenance commitment you own.
Read the frontmatter rules before you write any of it. The name must match the directory, and the description has to carry the vocabulary a user would actually type, because that is the discovery contract [12]. A skill with airtight enforcement and a vague description never runs.
