tov-guidelines — Stage 1 + 2 research skill (PMM spine)

Analyzes existing content (your website, blog, recent LinkedIn posts, sales-call transcripts) and produces voice rules that downstream content skills apply.

---

When to use

• When onboarding a new content writer / contractor — they need codified voice rules to write in your voice
• When LinkedIn or blog output stops sounding like the founder / brand
• Quarterly refresh

When NOT to use

• For visual brand identity (use /brand-kit — voice + visual are separate but both live in marketing/brand/)
• For one-off content checks against existing rules (write a custom gate skill that reads brand-voice.md)
• For positioning or messaging (use /positioning or /product-messaging)

How it works

1. Inputs: source content (one or more of: website URL, blog URL, LinkedIn handle, sales-call transcripts, founder writing samples).
2. Reads (Exa MCP for fresh data): the source content. Patterns extracted include: vocabulary preferences, sentence-length defaults, headline conventions, em-dash usage, list-vs-prose preference, contraction frequency, persona pronouns, signature opening / closing patterns.
3. Produces a two-phase output:
   • Phase 1 — Voice analysis (extracted patterns from source material with frequency scores): "uses em-dash with spaces 87% of the time across 30 LinkedIn posts; sentence case headings 100% of the time; banned-word violations 0 in last 10 posts"
   • Phase 2 — Voice guidelines (codified rules every content skill applies):
     • Rule N — {rule name}
     • The rule: 1-sentence statement
     • Why: 1-sentence justification
     • Violation pattern: what breaking this looks like
     • Fix template: how to fix
     • Exception: when the rule doesn't apply
4. Writes to marketing/brand/brand-voice.md (overwrites prior canonical; git history preserves prior versions). Marks status: locked once approved.

Invoke

/tov-guidelines https://yourdomain.com + analyze my last 10 LinkedIn posts

Or open-ended:

/tov-guidelines extract voice from these sales-call transcripts: [paths]

Example output

See marketing/brand/brand-voice.md for the PulseAnalytics seed. Notice the 5 rules each have rule statement + why + violation pattern + fix template + exception. The 5 rules cluster into two themes (prose hygiene + claim integrity).

Dependencies

• Reads from: source content (URL or files — provided as input); optionally marketing/icp/ICP.md (voice should match the ICP's vocabulary register)
• Reads via Exa MCP (optional): the source content if it lives on the web
• Writes to: marketing/brand/brand-voice.md (canonical)
• Downstream readers: any content skill (LinkedIn, blog, newsletter, email), /landing-page-copy, future gate skills

Customization

The 5-rule default in the PulseAnalytics example is operator-flavored. Yours may need more rules (large enterprise teams sometimes track 8-12) or fewer (founder voice often clusters into 3-4 rules). Use whatever rule count matches the patterns you can actually enforce.

Don't add a rule you can't enforce — every rule should map to a check a gate skill could run automatically. Vague rules ("be human") get ignored; specific rules ("no em-dashes without spaces") get enforced.

Where this fits in the Example 1 chain

Week 1: /tov-guidelines (parallel with competitor + ICP research) → voice rules locked
Week 2: /positioning + /product-messaging apply voice rules to output
Week 3+: every content / paid / outbound execution skill reads voice rules

Refresh cadence

Quarterly. Trigger sooner when the team finds a new buzzword pattern, customer flags off-brand phrasing, or a new persona is added (vocabulary may shift).