text-human — universal AI artifact humanizer

Before any processing the skill parses the argument, picks exactly ONE flow from context, and announces it: Flow: <name> — <one-line why>. Detection priority: explicit intent keywords (RU+EN) first, then path/extension, then content sniff. Only that flow file and the pattern sections it needs are loaded (lazy — not everything at once).

The first token is parsed as scope when it resolves to an existing path or a 7+ hex git hash; the rest becomes a custom prompt. Otherwise the whole input is a custom prompt (inline text to humanize or intent description). No args triggers AskUserQuestion as fallback. The custom prompt both selects/overrides the flow and adds custom rules at highest priority.

Before the first edit on any writing flow, the skill runs git status —porcelain over the resolved target paths only — there is no backup, so git is the only undo. Clean → proceeds. Dirty → one AskUserQuestion listing the dirty paths (proceed on them / process only the clean ones / abort); this replaces, not adds to, the flow-detection question. Not a git repo → the skill says so and requires an explicit go-ahead. Inline text and other non-file input skip this — nothing on disk changes.

Validated AI tells are removed per the tiered pattern catalog. HIGH-tier universals (chat scaffolding, self-attribution comments, prompt residue, unicode normalization in prose and code comments) act on single hits. Unicode inside a string literal, regex, or identifier is never rewritten — it is surfaced instead, since normalizing it could change program behavior. MED density-signals act only when several co-occur in a cluster. Behavior-changing items — hallucinated refs, fabricated tickets, try/except-everywhere — are surfaced for review, never auto-edited.

Human style is applied for the detected domain. HARD-OFF for code and API/formal docs. For docs, commits, social, and articles the injection is tuned by domain: burstiness weight, contraction frequency, stance, emoji policy. Global guard: never injects typos, errors, or fabricated references in any flow.

Commit or folder scope uses the mixed flow dispatcher: files split into blocks by complexity, classified as haiku (simple config/data/text) or sonnet (source with logic, tests, complex SQL). All Task calls fire in one message for true parallelism. Each sub-agent receives its block, the flow rules, and any custom instructions prepended at highest priority. The skill substitutes the resolved absolute path of the rules directory into every sub-agent prompt before spawning — subagents get no plugin-root variable of their own, so an unexpanded one would leave them humanizing against no rules at all. Commit mode is current-worktree: the hash only selects which file names changed, then edits land on today’s content at those paths, not the historical blob — the same clean-tree gate runs again here, scoped to the commit’s paths, before any block agent spawns.

Results merge into a unified report: detected flow, scope, files/blocks processed, haiku/sonnet split, per-file strip and inject counts, a surfaced-for-review section listing behavior-changing items, and totals. Files are edited in place — use git to revert.