Brand Voice Extractor

Brand Voice Extractor

Fetch a brand's published pages, strip them to clean body text, compute deterministic style metrics (readability, sentence/paragraph length, POV pronoun mix, header/list density), then you synthesize the Brand Voice Profile from that evidence. The scripts do only I/O and measurement — all voice/tone judgement is yours.

When to use

• Before writing content/outreach/campaigns that must sound like an existing brand.
• "Extract brand voice for [company]" / "match [client]'s tone."
• Onboarding a new client with no voice/style guide.
• Upstream feeder for messaging, positioning, or content-generation skills.

How to run

Scripts are Python 3 stdlib-only (no install). The JS fallback needs Playwright.

1. (Optional) Discover candidate URLs from a blog/site root

When you only have a root URL, enumerate + classify candidates (sitemap first, then crawl):

python3 ${SKILL_DIR}/scripts/discover_urls.py \
  --root https://acme.com/blog \
  --max 200 \
  --output ${WORKSPACE}/candidates.json

Then you select a diverse 10–20 sample from candidates.json — favor blog posts (~8–10: how-to, opinion, product), 2–3 landing pages, 2–3 case studies, 1–2 comparison pages; mix recent + older to catch voice evolution. Blog posts carry the strongest voice signal; landing pages are formulaic — weight blogs.

2. Fetch + measure the chosen pages

python3 ${SKILL_DIR}/scripts/fetch_content.py \
  --company "Acme" \
  --urls https://acme.com/blog/a https://acme.com/blog/b ... \
  --num-pages 15 \
  --output ${WORKSPACE}/corpus.json

corpus.json has per-page {url, title, word_count, text, metrics{...}} plus a corpus_metrics roll-up. Pages that come back as thin text (errors[].error mentions "JS-rendered") should be re-fetched with the Playwright fallback:

# one-time: npx playwright install chromium   (and: npm --prefix ${SKILL_DIR}/scripts install)
node ${SKILL_DIR}/scripts/fetch_content_js.mjs \
  --company "Acme" \
  --urls https://acme.com/blog/js-page ... \
  --output ${WORKSPACE}/corpus-js.json

3. Synthesize the Brand Voice Profile (you, the agent)

Read corpus.json (+ any corpus-js.json) and write a Markdown Brand Voice Profile:

1. Voice summary — 2–3 sentences capturing the overall voice.
2. Tone profile table — formality, emotional register, authority stance, humor, directness; attach a real quote from the corpus as evidence per dimension.
3. Language & vocabulary — reading level (use flesch_* from metrics), jargon level, power words, signature phrases, words loved/avoided.
4. Structure & formatting — sentence/paragraph length (from metrics), header/list/ emphasis density (from the h/li/ul_ol/strong_em counts), CTA + callout habits.
5. Audience / persona — target reader, knowledge assumptions, POV (use pov_totals: we/our vs I/me vs you to infer first-person-plural vs second-person address).
6. Do / Don't writing guidelines + 2–3 "how to match it" sample sentences on a neutral topic.
7. Inconsistencies — flag content types whose metrics/tone diverge sharply (possible multiple voice modes or ghost-writers) rather than averaging them away.

Outputs

• corpus.json (and optional corpus-js.json) — deterministic fetch + metrics artifact.
• Brand Voice Profile Markdown (brand-voice-profile-<company>.md) — your synthesis, returned as the result and saved to the workspace; for team use, post to the Agent Teams channel rather than an invented path.

Credentials / env

• Required: none. Page fetch is keyless (public HTTP); the keyless path is the default.
• Optional (paid upgrades, each with a keyless fallback):
  • APIFY_API_TOKEN — if set → route a too-hostile blog page through an Apify article-extractor actor. If not set → default keyless path: fetch_content.py (stdlib), then the Playwright fallback fetch_content_js.mjs. Apify is the last resort, never required.
  • DATAFORSEO_LOGIN/DATAFORSEO_PASSWORD (or SERPER_API_KEY/SEARCHAPI_API_KEY) — if set → use SERP to discover the brand's best content URLs (better recall on sites with no/partial sitemap). If not set → default keyless discover_urls.py (sitemap.xml → same-host crawl).
• Voice/tone synthesis is done by you (the host agent) — no LLM key is consumed by the scripts. See env.optional.

Notes & edge cases

• 15 pages is the sweet spot: <10 misses variation, >25 adds cost without signal.
• fetch_content.py flags thin pages (< --min-words, default 80) as likely JS-rendered — escalate those to fetch_content_js.mjs before giving up.
• discover_urls.py tries sitemap.xml (incl. sitemap index) first, falls back to crawling same-host links on the root; it drops asset/tag/author/pagination URLs.
• All scripts back off on HTTP 429/503 and space requests (0.3s) to one domain. For heavy fetching from a single host behind IP limits, route through a proxy at the platform layer and/or reduce --num-pages.
• Metrics are deterministic facts (readability, lengths, pronoun counts) — score the voice from those real numbers, don't guess.