An audit that ends with “this reads well” is not an audit. A client pays for a verdict they can check, and a model that produces one has to be constrained before it writes a word. That constraint is a scorecard, and the five rules below decided what went into mine.
Code: RankWise, MIT licensed.
RankWise 1.5.1 scores content against 49 factors in nine languages, with a Python validator for the measurable part. It is a SKILL.md and a folder of references, with no API keys and no dependencies, so it loads into any capable model.
Rule 1: a factor has to be checkable by a second person
Every one of the 49 factors is written so that two reviewers reading the same page reach the same answer. “Compelling introduction” fails that test. “Focus keyword appears in the first ten percent of the text” passes it, because the second reviewer can count.
The rule sounds obvious and it is the hardest to keep. Half of the advice in a typical SEO checklist is a taste claim wearing a number.
Rule 2: weights come before the total
Each factor carries a multiplier: CRITICAL ×3, HIGH ×2, MEDIUM ×1, LOW ×0.5. One failure on a CRITICAL factor caps the grade at B, whatever the sum says.
A plain percentage hides the order of repairs. A page at 92 with no canonical tag is not a better page than one at 78 with everything in place, and the capped grade says so without a paragraph of explanation. It also ends the meeting about which fix goes first.
Rule 3: a factor that does not apply leaves the denominator
A page without images cannot fail an alt-text check, and a page without a publication date should not be punished for freshness. Such factors are removed from the denominator instead of scoring zero.
Without this rule a score compares pages of different shapes, and the comparison is meaningless in exactly the case where a client wants it: a service page against a blog post.
Rule 4: the machine counts what the machine can count
Sixteen factors are computed deterministically: word count, paragraph metrics, readability, passive voice share, transition words, sentence variety, repeated openings, keyword density and placement, title and description length, power words.
The other 33 need a judgement about meaning, and the report labels them as judgement. That split is the honest part of the design. A number a script reproduces can be argued with evidence; a number invented by an LLM is an opinion with a decimal point.
Rule 5: language is a parameter, not a translation
The formulas change with the language: Flesch-Kincaid for English, Wiener Sachtextformel for German, Fernández-Huerta for Spanish, Gulpease for Italian, LIX for Ukrainian, FOG-PL for Polish. Title and description length targets move too, because the same idea takes a different number of characters in each language.
Passive voice patterns and power words are language-specific as well, and a list translated from English produces false positives in Russian within a minute. Nine languages are supported because clients are not all English-speaking, and an audit in English on a Ukrainian page is worse than no audit.
What the 49 factors do not cover
Page speed, Core Web Vitals, server configuration, SSL and cookie compliance are outside the scorecard. Those are ranking inputs, and a content skill that claims them would be describing a tool that does not exist.
I would rather state the boundary in the first screen of the readme than let a client discover it after paying for an audit that never looked at the server.
Why a skill instead of a plugin
A plugin needs a home, a version of WordPress, a maintenance window. A skill is a text file that loads into whatever model the team uses, including the free tier of something that runs on a laptop.
The validator is optional: without the Python package the sixteenth factor returns N/A, and the other fifteen still compute. That is the same design decision as the queue drivers in another project of mine: the project works on the smallest environment it will meet, and grows when the environment does.
What it teaches
Write the standard before the tool. A scorecard that two people can argue about is worth more than a model that sounds confident, and the argument is where the client’s money actually goes.
The same idea runs through the AI copywriting skill: the technique is documented, the assumptions are written down, and the output can be checked. Content angle is the step before this one, where the piece is aimed at a reader instead of a keyword.
RankWise is MIT, installs with one Python command, and works standalone as a system prompt.