Brand Voice Discover

Brand Voice Discovery

Gather every available signal about how the brand sounds — what Bolta already stores, what the brand has already published, and anything the user uploads — then output a structured discovery report that brand-voice-generate turns into a guideline. Do not invent voice; only surface, rank, and reconcile evidence.

When to use

The starting point of the brand-voice loop, or any time the user wants to understand their current voice signal before committing to a guideline. Run this first when no session guideline exists.

Tools this skill uses

Two more sources need no tool call: the brand's website (you can read public pages directly since ChatGPT has internet access) and what ChatGPT already knows about a recognizable brand. Both are supporting evidence — always confirm against the brand-owned Bolta signal, never let outside knowledge override what the brand actually publishes. Heavy parsing of uploads and posts is done inline using the analysis behaviors below.

Prerequisites

• workspace_id — resolve once via list-workspaces and reuse it for every call. Auth is automatic via the Bolta connector's OAuth grant — never ask for an API key. Default any new content to Draft; confirm before publish/delete.
• Optional user uploads: brand docs (PDF/PPTX/DOCX/MD/TXT), sales/interview transcripts, or pasted sample copy. Parse these inline: brand documents and transcripts via the signal-distiller behavior (Step 5), and the brand's own published posts via the voice-archaeologist behavior (Step 4). A brand website URL is a strong optional input too (Step 3).

Workflow

1. Resolve the workspace

Call list-workspaces; use the active workspace's id. If several exist, confirm which one with the user. Never guess a UUID.

2. Pull existing Bolta signal

In parallel where possible:

• list-business-dna → get-business-dna for each record (industry, audience, positioning). If the user spots something stale or wrong in a record while reviewing it ("that tagline is old", "we moved — the city is wrong", "our colors are navy and sky blue now"), fix it conversationally with update-business-dna(workspace_id, dna_id, fields={…changed keys only…}) — it's a safe partial merge, omitted fields are preserved, so a one-field correction can't wipe the rest. Confirm the change before writing.
• list-voice-profiles → get-voice-profile for each profile (stored tone, dos/donts, exemplars).
• get-voice-context — the compiled context Bolta's writer already uses, if any.
• list-accounts — which platforms are connected (context for where the voice lives). This is authoritative brand-owned signal; weight it highest.

3. Pull website + external signal

If the user gives (or you can infer) the brand's website, use it — it's the fastest cold-start source and often the clearest statement of positioning and voice:

• Read the public site directly (homepage, about, product pages) for headline voice, value props, and terminology; and/or call extract-business-dna with the url to structure it into a Business DNA record you can reference later. Pass set_as_default=false — the param defaults to true and silently demotes the workspace's existing default Business DNA. Only set it true (or omit it) when the user explicitly wants the extracted record to become the default, and confirm first if a default already exists.
• If the brand is recognizable, you may draw on what ChatGPT already knows about it — but treat this as a hypothesis to confirm, never as fact. State that it's external knowledge. Rank this below brand-owned Bolta signal and published posts: the website says how the brand wants to sound; the posts show how it actually sounds. When they disagree, the posts win and the gap becomes an open question.

4. Pull published-content evidence

Call list-workspace-posts with status=published and include_metrics=true — each post then carries per-platform engagement (views, likes, comments, reposts) alongside its text. The default page is 50 posts (max 100); page through with limit/page so the evidence isn't silently truncated to the first page. Published posts are behavioral voice evidence — how the brand actually talks in public. Excavate the brand's lived voice from Bolta's own evidence (the voice-archaeologist behavior): pull the published posts with their per-post metrics, and the review signal (list-reviews, list-recurring-reviews) — what the team approved untouched vs. edited vs. rejected. Extract voice attributes each backed by a quoted real post, ranked by consistency; mark attributes that correlate with higher engagement as "proven" not just "present"; treat systematic edits (always cutting hype, always adding a number) as implicit rules and rejected phrasings as "We Are Not" candidates. Redact customer names/PII from quotes. Weak evidence = 1 post; strong = a pattern across 8+.

5. Classify and parse uploads

For each user-provided asset, distill it inline using the signal-distiller behavior: turn uploaded materials (decks, one-pagers, brand books, transcripts, site copy) into Bolta's structured models — Business DNA {industry, audience, positioning, offer} and Voice Profile fields {tone dials OR a writing_sample excerpt, tone_of_voice[], dos[], donts[], target_audience, terminology: must-use/preferred/avoid/never}. Attribute every field to its source; leave ungrounded fields empty rather than inventing; prefer a strong on-brand writing_sample over guessing precise tone dials when material is qualitative. Where uploaded materials disagree with the brand's shipped posts, flag it as an open question rather than silently picking one. Transcripts (sales calls, founder interviews) often carry the truest, unscripted voice signal. Keep only the structured extractions.

6. Reconcile and rank

Merge every source into one picture. Rank sources by reliability (brand-owned DNA/Voice Profiles > published posts > brand documents > transcripts > pasted samples), noting recency and volume. Flag conflicts explicitly (e.g. DNA says "formal", published posts read casual).

7. Output the discovery report

Produce a structured report:

• Ranked sources — each with type, reliability weight, recency, and volume.
• Extracted attributes — candidate "We Are" traits, each with the evidence behind it.
• Terminology signal — words the brand uses / avoids.
• Tone-by-context observations — how register shifts by platform/topic.
• Conflicts — where sources disagree, with both readings.
• Gaps — attributes with thin or no evidence.
• Open questions — what brand-voice-generate must resolve, each with a suggested default.

Failure handling

• No workspace / permission error → report the missing scope (get-my-capabilities) and stop.
• No Business DNA, Voice Profiles, or voice context yet → say so plainly; lean on published posts and uploads. A brand-new workspace with no posts and no uploads = not enough signal; ask the user for sample content rather than fabricating a voice.
• No published posts → note the gap; discovery still proceeds from DNA/uploads, but flag that behavioral evidence is missing so confidence will be lower downstream.
• Upload unreadable → report which asset failed and continue with the rest.

Example

User: "Analyze our brand voice — here's our pitch deck and last quarter's posts are in Bolta."

1. list-workspaces → workspace_id.
2. list-business-dna/get-business-dna → SaaS, ops leaders, "developer-first" positioning; list-voice-profiles empty; get-voice-context empty; list-accounts → LinkedIn + X.
3. list-workspace-posts(status=published, include_metrics=true) → 40 posts → voice-archaeologist pass → "direct, proof-driven, dry humor; avoids hype; LinkedIn more formal than X."
4. Pitch deck → signal-distiller pass → terminology ("platform" not "tool"), 3 messaging pillars.
5. Reconcile: deck says "visionary", posts read "pragmatic" → conflict flagged.
6. Report: 5 ranked sources, 6 candidate attributes with evidence, 1 conflict, 2 gaps, 3 open questions → hand to brand-voice-generate.