Sage — The Subject Matter Expert Builder
A pipeline that turns video corpora and practitioner interviews into callable Subject Matter Expert agents. The meta-build that produces the tools I use to build everything else.
The Problem
I wanted to ask follow-up questions of practitioners who weren’t in the room. Read a great blog post on local SEO? Watch a 50-video YouTube course on negative-keyword strategy? Those experts have answers I’ll never get unless I either fly to them or build something that captures their thinking.
So I built Sage — a system that takes a corpus of video content and practitioner work and turns it into a Subject Matter Expert agent I can talk to. Defended claims. Confidence tags. Disputed-vs-verified-vs-confirmed labels on every answer.
The Architecture
Five stages, each one separately auditable:
- Transcribe — Whisper-based transcription pipeline, deployed on a DigitalOcean droplet behind
transcribe.reelygood.online. Handles long-form video, multi-speaker audio, accents, technical jargon. - Extract — LLM-powered claim extraction from raw transcripts. Each claim is timestamped, sourced, and tagged with a topic.
- Verify — adversarial review pass: every claim is checked against the original transcript and against external sources. Claims earn one of three labels: Verified, Confirmed, or Disputed.
- Synthesize — claims are clustered into a canonical Q&A canon. Questions the practitioner answers repeatedly become canon questions. Conflicts between practitioners are flagged, not hidden.
- Materialize — the SME ships as a callable Claude subagent with its canon embedded. You ask it a question; it answers with confidence-tagged citations to the original sources.
Built SMEs to date: sme-personal-brand-video (36 videos, 106 claims), sme-lead-scoring-hubspot (389 claims, 156 canon questions), sme-negative-keywords-google-ads, sme-local-seo (49 videos, anchored on Whitespark + Sterling Sky canon).
Key Decisions
Why confidence tags and not a single “right answer.” Practitioners disagree. The system that pretends they don’t is lying. Verified / Confirmed / Disputed is honest about what the corpus actually contains — and lets the user weigh contested claims themselves.
Why an adversarial verification pass. An LLM that extracts claims and an LLM that approves its own claims is the generator-verifier failure mode I’ve written about. A separate verifier — different model, different prompt, no memory of the extraction — catches the fabrications and the half-truths.
Why canon questions, not raw transcripts. People don’t ask “what did the speaker say at 14:32?” They ask “should I add this negative keyword?” The canon is organized around the questions a practitioner would actually ask in the field.
Why ship as a Claude subagent, not a chatbot UI. Subagents are composable. I can invoke the local-SEO SME from another agent, from a script, from a workflow. A chatbot UI is a dead end.
What I Learned
-
The hardest part of building an expert is admitting what they don’t know. The temptation is to make the SME answer everything. The right move is to teach it to say “this isn’t in the corpus” — and mean it.
-
Verification is the moat. Extraction is commodity. Anyone can chunk transcripts and run an LLM over them. The thing that makes Sage actually useful is the second pass that throws out the hallucinations the first pass loved.
-
Built-in disagreement is a feature. Once you stop hiding contradictions between practitioners and start surfacing them — “Joy Hawkins says X; Whitespark canon says Y; here’s why they disagree” — you’ve built something more useful than any single practitioner’s blog.
Status
Operating as of May 2026. Multiple SMEs in active use as part of my own AI workflow. The pipeline itself lives across ~/projects/tools/sage/ and ~/.claude/skills/.