Two modes sharing one artifact: Learn deep-crawls a website and produces a self-contained style guide; Apply rewrites documents in that guide's voice so a whole batch reads as one author. The guide file is the only thing that crosses sessions — everything else is working state.

Bundled files (load on demand):

Output: (host without ). Working state: (, , ) — resumable, deletable after the guide lands.

Learn mode drives a real browser through whatever automation the agent has — Playwright MCP ( bridge to the user's live Chrome, or a spawned browser), or an equivalent browser tool; without one, say so and stop (a plain HTTP fetch tool cannot render JS-heavy sites and silently misses content — do not degrade to it without telling the user).

With the Playwright MCP bridge, run the target gate first (the procedure is in , shipped with ): ask which bridge when the session exposes more than one, probe which browser and profile answered, and confirm with the user before the crawl starts. Wrong browser or no bridge → ask for that browser's from the extension's status page, set it in the MCP entry, restart, re-run the gate. Then list tabs (a lone means the bridge is not attached — stop and have the user fix it; no extension installed at all → the install for Chrome or Edge is https://chromewebstore.google.com/detail/playwright-extension/mmlmfjhmonkocbjadbfplnigmagldckm); never touch the bridge's own tab; open ONE working tab and reuse it. Warn the user the browser is busy while the crawl runs. Dismiss cookie/consent banners once on the first page — they pollute extracted text.

All bookkeeping lives in the bundled ingester:

node <skill-dir>/scripts/crawl-ingest.mjs --dump dump-<host>.json --dir style-crawl/<host> [--origin <url>]

It ingests a dump of fetched pages and prints one JSON line: (up to 8 URLs), counters, coverage %, . No → just re-read state; that is also how a crawl RESUMES after interruption. First run needs .

The loop is two tool calls per batch of 8 pages:

html = html.replace(/<\/(p|div|h[1-6]|li|section|article|ul|ol|blockquote|tr)>/gi, '\n</$1>');

Fast-path validity check: extract the FIRST page twice — live (navigated tab) and via — and compare. Empty or much shorter fetch text → the site is JS-rendered: fall back to per-page navigation + a settle wait + the same extraction in the live DOM, still dumping to the file and ingesting with the same script (single-page dumps are fine).

Rules:

Serial = templated pages whose count can run to thousands (blog posts, products, glossary terms, tags). The script samples them (default quota 18 per pattern) instead of exhausting them: auto-serial path patterns plus automatic promotion when ≥8 URLs share a template. fires at ≥80% coverage of NON-serial pages with all quotas met, or an empty frontier. Safety cap 300 pages: on , STOP and tell the user the real coverage — never present a capped crawl as full.

Split into batches of ~15 files; spawn one subagent per batch, all concurrently. Each reads its files and writes STYLE observations (not content summaries) to with fixed sections: Lexicon / Voice & POV / Rhythm / Structure / Formatting / Genre notes / Golden-sample candidates (3–5 verbatim excerpts ≤120 words with source file and why), returning only a 5-line summary. Resource preflight (before fan-out): cap concurrency at , ≈ 0.7 GB for these read-only agents; go serial if CPU load > 85% or free RAM < 2×per_agent; recompute before each wave; where the runtime caps sub-agent concurrency itself, defer to it.

One agent (or the main context) reads all , reconciles (majority wins; genre differences become sub-profiles, not contradictions), and writes with exactly these sections: Voice profile · Tone rules (Do/Don't) · Lexicon · Rhythm & syntax · Structure (with the site's invariant CTA strings quoted verbatim) · Formatting habits · Genre notes · Samples · Rewrite instructions. writes the same section set for an author's own voice, so either file can be handed to Apply mode or to — keep the names exactly as listed rather than improving them. The guide must be self-contained — Apply sessions see only this file. See for the target shape and depth.

Two sample policies — pick by the guide's destination, ask when unclear:

Inputs: a style-guide path + a target (file or folder). Read the guide FIRST, fully — its Golden samples anchor the tone; its Rewrite instructions override defaults below.

Single file → one rewrite. Folder → glob prose-bearing sources ( — component files: copy strings only), skipping , build output, lockfiles, pure code/config. List the set first when >20 files.

Target inside a git repo ( exits 0) → ask which mode, unless the user already named one:

Non-repo target → mirror folder by default; in-place only on explicit request.

One subagent per file, spawned in parallel. Each subagent gets: the style-guide path, one source path, one output path, the file mode (markdown/html/component), and the FULL text of — identical guide + identical contract per file is what keeps one authorial voice across the batch. Never relay a summary of the guide; each subagent reads the guide file itself. A failed file gets one retry, then is reported — never silently dropped. Resource preflight (before fan-out): cap concurrency at , ≈ 0.7 GB for these read/write agents; go serial if CPU load > 85% or free RAM < 2×per_agent; recompute before each wave; where the runtime caps sub-agent concurrency itself, defer to it.

After all rewrites land (2+ files), run ONE consistency-pass subagent over the whole output set (for >15 files: first/last 3 paragraphs plus a middle excerpt each): find cross-document drift — lexicon used in one file but violated in another, tone shifts, inconsistent heading/CTA patterns — fix findings directly with edits, return the list. Report that list to the user; it is the evidence the batch reads as one author.