guideline-generation

Generates LLM-ready brand voice guidelines from any combination of source materials, including a structured discovery report, brand documents (PDF, PPTX, DOCX, MD, TXT), conversation transcripts from tools like Gong, Granola, or Notion meeting notes, and direct user input. When a discovery report is available it is used as the primary, pre-triaged input; otherwise the skill can check for an existing report from a prior discover-brand run or consult a local brand-voice file for known material locations before suggesting that discovery be run first.

Processing delegates heavy parsing to specialized agents: a document-analysis agent extracts voice attributes, messaging themes, terminology, tone guidance, and examples, while a conversation-analysis agent recognizes implicit voice attributes, successful language patterns, tone by context, and anti-patterns. Findings are synthesized into a unified document built around several anchors: a "We Are / We Are Not" identity table, a distinction between voice constants (personality and values that stay fixed) and tone flexes (formality, energy, and technical depth that adapt by context), and a tone-by-context matrix covering scenarios such as cold outreach, enterprise proposals, and social media. Each section receives a confidence score (high, medium, or low) based on how many corroborating sources support it, and every unresolved ambiguity becomes an open question that must include an agent recommendation so the user can confirm or override rather than hit a dead end.

Before presentation, a quality-assurance agent verifies that major sections are populated, that there are at least three evidenced voice attributes, that the identity table has four or more rows, that the tone matrix covers at least three contexts, that confidence scores and source attribution are present, and that no PII is exposed. Privacy constraints are enforced throughout, redacting customer names and contact details and optionally anonymizing company names in transcript excerpts. Completed guidelines are saved to a .claude/brand-voice-guidelines.md file inside the user's working folder, archiving any prior version by date, with the path resolved relative to the user's working folder rather than the current directory; downstream enforcement tooling can then pick them up automatically.