Six skills.
One job each.

Does. Reads the whole draft, applies structural edits first and vocabulary second, then runs a preservation pass, then scores the result against a vendored humanizer rubric. A score under 90 sends the passages that lost points back for a rewrite and a rescore, at most three times.

Refuses. Em-dashes anywhere, in any form. Em-dash lists, single-noun colon-reveals, and semicolons between independent clauses. “Simply”, “just”, “obviously”, “of course”, and “clearly”. Adding hedges, asides, or jokes to look human. Claiming a detector result. Restructuring the argument, beyond a flagged reorder inside a paragraph.

Protects. Accurate hedges already present, the confessional or first-person opener, grounded asides, technical terms that are the right word, and a fragment sequence that lands a build-up. Where a deduction lands on protected text, the skill keeps the text and records kept (−N): reason.

Does. Substitutes terms from a Microsoft Writing Style Guide list filtered for developer content: blocklist for blacklist, allowlist for whitelist, primary and replica for master and slave, soundness check for sanity check. Fixes disability metaphors used figuratively, gendered defaults, “email”, “internet”, “website”, setup as a noun and set up as a verb, the Oxford comma, numerals from 10 up, and sentence-case headings. Adds H2 headings when a draft has none and needs them.

Refuses. Changes to sentence structure, tone, hedges, or inline code. It also refuses scope creep by name: clarity edits belong to humanize, and running both in the wrong order makes the second one undo the first.

Judgment calls are reported. A term that is being discussed rather than used (“the old term was blacklist”) stays, and the changelog says why.

Does. Applies the four governing principles of ISO 24495-1:2023. Readers get what they need, can understand it, can find it, and can use it. Kills nominalisations, fixes ambiguous references, orders information the way the reader will use it, and asks for a reader test where one is possible. When reviewing, it names the principle, quotes the passage, explains the failure in reader terms, and gives a revision.

Refuses. To work without a stated audience and purpose. It asks first, and it says so: you cannot apply Principle 1 without knowing who the readers are and what they need to do. It also refuses to dumb down or to strip information the reader needs to act correctly.

Does. Applies ASD-STE100 to procedures, descriptions, and safety text. Procedural sentences cap at 20 words and descriptive ones at 25. One instruction per sentence. Active voice, simple tenses, no ambiguous participles, at most three nouns in a cluster, one topic per paragraph, and abbreviations defined at first use.

Refuses. To assert that a word is STE-approved when it cannot check. The full dictionary is not loaded, so an uncertain word is flagged needs verification with a likely substitute. It keeps technical names and technical verbs regardless of the word list, and it will not shorten a safety warning below the length that keeps it unambiguous.

Scope. Do not run it on narrative or marketing copy unless you asked for that. It is deliberately austere, and it flattens voice by design.

Does. Works down five layers in order: truth and scope, substance, development, sentences, craft. Reports a piece-level diagnosis, then passage findings, each carrying one of four verdicts: keep, revise, ask-author, or cut. Runs five tests before naming a finding: substance, flatten, relation, specificity, and triad. Closes by collecting the open ask-author questions into one list.

Refuses. To produce a replacement draft, to edit the file, or to stage anything. “This is ready” is a valid outcome, and the skill says so explicitly rather than manufacturing findings.

Ask-author is the load-bearing verdict. It is the alternative to inventing the missing detail, which the skill names as the easiest way to damage a draft while believing you improved it.

Why it exists. The profile’s own rule, as the skill quotes it: the voice is portable and the thinking is not. A model can reproduce cadence from a profile. It cannot reproduce what the author saw, measured, decided, or changed their mind about. So the interview comes first, and drafting waits.

The hard gate. No opening paragraph, no outline presented as settled, no rough angle to react to. Ask, then wait. For a piece that needs no personal material, the skill says it is the wrong tool and hands off rather than running the ceremony.

Shane-specific by construction. It names him, opens with a prompt about what happened that made him want to write the piece, sweeps his vault for prior material through the Obsidian CLI, and reports a provenance block naming author material, model contribution, and open items. It also assumes a publishing workflow that refuses to ship while a [TK] is unresolved.

What travels. The procedure, if you rewrite the questions for your own material and your own sources. The skill as shipped does not, and it should not: the value is in interrogating a specific author, not in a generic interview ritual.

Input, verbatim:

Humanize this: "The approach is more specific. More actionable. Closer to the real problem."

The case expects the three fragments to merge into one sentence with all three qualities still present. Its assertions are that the fragments are joined into a single sentence rather than left standalone, and that more specific, more actionable, and closer to the real problem all survive.

Input, verbatim:

Humanize this: "The build finished in 40 minutes. The team shipped it the same day."

This one fails if the pass over-corrects. The sentences must stay committed, no hedge may be introduced (the assertion names so far, with any certainty, about, roughly, may, might, could, perhaps, seems, appears, and tends), “not only” must not appear, and the first sentence must remain a plain assertion rather than being softened to imply uncertainty.