TOV guidelines

Extract actionable editorial tone of voice guidelines from a company's existing content. Two-phase workflow: Phase 1 analyzes website content to surface evidence-based voice patterns; Phase 2 codifies those patterns into guidelines that downstream content skills consume. User review gate sits between the two phases.

When to run

• User says "tone of voice", "brand voice", "voice guidelines", "writing style guide", or asks to extract voice from a URL.
• Starting a new content program for a client and downstream skills (linkedin-content, landing-page-copy, product-messaging, outreach-emails) need a voice contract.
• Existing TOV is stale (>6 months) or content has drifted from documented voice.

Don't run when: user wants visual brand identity (use brand-kit), product messaging (use product-messaging), to write content directly (use the content skill with TOV as input), or just a quick voice description (answer directly).

Inputs

Input	Required	Source
Company URL	yes	User
Founder context (transcript, notes, voice preferences)	optional	User / company-context
Existing brand guidelines (to update)	optional	User
Target audience	optional	icp-research / user

Validate before proceeding: URL is accessible, site has homepage + 5-10 pages of content. If thin, flag sample-size limitation in output.

Research substrate (Exa): primary tool web_search_exa for founder voice harvesting via site:linkedin.com / site:twitter.com filters. Per .claude/rules/exa-protocol.md. Citation: [VERIFIED: exa_search, {url}, accessed {YYYY-MM-DD}]. Quality gate: ≥3 sources per major claim, ≥50% [VERIFIED].

Steps

Two phases. Phase 1 outputs tov-analysis.md. User reviews. Phase 2 outputs tov-guidelines.md.

Phase 1 — Analysis

1. Discover site structure — fetch homepage, extract internal links, filter same-domain only, prioritize /about, /manifesto, /values, /blog/*, /case-studies/*, /pricing, /faq.
2. Scrape 15-20 pages in priority order: homepage → about/values → blog (3-5) → case studies (2-3) → pricing → FAQ → landing pages.
3. Extract sentence-level patterns — average length, person (1st/3rd), question frequency, imperative usage.
4. Extract paragraph-level patterns — average length, openings, transitions, evidence placement.
5. Extract word-level patterns — company vocabulary, customer/industry vocabulary, banned/avoided words, modifier frequency.
6. Extract structural patterns — headers (sentence vs title case), CTA placement and phrasing, proof stacking, section organization.
7. Score frequency — High (80%+ of pages), Medium (40-79%), Low (<40%), Conflict (contradictory).
8. Build content-type voice mapping — table of person/formality/CTA per page type (homepage, blog, case study, pricing).
9. Generate voice-in-action examples — generic → on-brand transformations, drawn from actual scraped text (never invented).
10. Identify inconsistencies — flag conflicts (e.g., homepage uses "I", pricing uses "we") for Phase 2 resolution.
11. Document gaps — what couldn't be determined; suggest founder interview questions.
12. Write tov-analysis.md using the premium reference (44-section canonical scaffold; compact alternative in same file).
13. Present to user — Review gate. User must confirm patterns, correct misidentifications, and answer gap questions before Phase 2.

Phase 2 — Generation

14. Incorporate user feedback — apply corrections, resolve inconsistencies, fill gaps with founder answers.
15. Generate tov-guidelines.md using the premium reference — primary reader, tone attributes with before/after, pattern library with LLM guidance, vocabulary lists, structure templates, anti-patterns. Always include an AI-speak anti-patterns section — select ≥5 relevant rows per the premium reference and embed them inline with the client's own examples, stripping row numbers and internal paths so the client's doc stands alone. 15a. Append the 'We Are / We Are Not' table per the premium reference Template 1. Source every row from Phase 1 evidence; mark confidence per row; overall confidence is the worst of any row. Minimum 5 rows, maximum 8. 15b. Append the tone-by-context matrix per Template 2. Cover the 7 default contexts (cold outbound, nurture, objection-handling, sales-call follow-up, support reply, case study, social post). Add or split rows for client-specific contexts. 15c. Append the open-questions list per Template 3. Minimum 3 questions, each with a "Why it matters" line naming the downstream unblock.
16. Add source attribution — every guideline traces to a source URL; frequency scores carried forward; unresolved gaps marked "TBD — requires founder input".
17. Self-evaluate per the premium reference (completeness, evidence, guardrails, self-roast). Surface improvement prompts to user when checks fail.
18. Save approved output as reference example if user explicitly approves — see the premium reference.

What good looks like

Evaluations

A TOV output passes when:

• 15+ pages scraped, all pattern categories (sentence, paragraph, word, structural) have findings.
• Frequency scores based on actual counts (e.g., "5 of 6 pages, 83%"), not "usually" or "often".
• Every guideline traces to a source URL.
• All examples are from actual scraped text (no invented illustrative examples).
• Inconsistencies explicitly flagged, not papered over; gaps documented with founder questions.
• No adjective-only descriptions ("friendly", "professional"); no vague guidelines ("use conversational language").
• All 6 critical questions answered: who reads / what tone sounds like / patterns repeat vs vary / words used / structure / what to refuse.