- name:
- advocate-voice-tuning
- description:
- Use this skill when AI-drafted Reddit comments don't sound like the person they're supposed to be — "the drafts feel off-brand," "make the advocate sound like me," "the AI gets our product details wrong." Two levers: teach the voice by rewriting sample drafts into your natural phrasing, and feed a tight knowledge base so drafts get the facts right.
Use when drafted comments miss the voice or the facts. Produces trained voice guidelines plus a focused knowledge base — the two inputs that turn generic AI comments into ones a teammate would sign.
Teach the voice by rewriting
Take a handful of real drafts (five is enough) and rewrite each the way you'd actually say it — or approve the ones that already sound right. Even 2–3 rewrites carry useful signal. The deltas between draft and rewrite become explicit voice guidelines: vocabulary ("uses" vs "leverages"), sentence rhythm, hedging style ("in my experience…" vs flat statements), how often and how directly affiliation gets disclosed, and comment structure (empathy-led vs question-led vs statement-led).
Do it well:
- rewrite instinctively — first-thought phrasing is the authentic signal, polished phrasing isn't
- rewrite from the Reddit post, not from the AI's draft
- include your natural disclosure style in the rewrites
- pick samples from varied subreddits so the guidelines generalize
Review the extracted guidelines afterward and hand-edit what the extraction missed. Sessions accumulate — rerun after off-brand batches, repositioning, or when spinning up a new persona.
Feed the facts
Voice without facts still produces wrong comments. Attach a knowledge base per campaign: product overview and differentiators, feature specifics, real use cases, pricing, FAQs, tone examples. The drafting engine retrieves only the chunks relevant to each thread, so focus beats volume — a tight 5-page knowledge base on one product outperforms a 100-page dump. Skip raw code, image-dependent decks, and internal ops docs (images aren't indexed — only surrounding text). Refresh it after every major release; stale knowledge produces confidently outdated drafts. One campaign per product keeps retrieval clean.
What good looks like
A tuned advocate's drafts are indistinguishable from the person's own comments — same rhythm, same hedging, same disclosure habits — and factually current. The overlooked failure: tuning voice once and never refreshing the knowledge base, so drafts sound right but describe last quarter's product. Success: the reviewer approves drafts with fewer and fewer edits over time.
MUST review and edit extracted voice guidelines rather than trusting extraction blindly. MUST refresh the knowledge base after major product changes. NEVER pad the knowledge base with content that doesn't help answer a buyer's thread.