How We Built Avery Young’s Authority Package Live on Zoom

On July 2, Avery Young-Nulf of HighLevel watched me tell Claude, live on our Zoom call, to build his personal authority audit, a full Authority Engine program strategy, and a microsite honoring him. By that evening the agent had shipped two diagrammed strategy PDFs, published dennisyu.com/avery, and staged the outreach email with both PDFs attached. This is the meta article documenting exactly how.

Watch the Assignment Happen Live

Avery reached out cold on June 24. Eight days later we were on Zoom architecting what became the Authority Engine: collect everything an entrepreneur is credibly great at, score it, and feed it into their landing pages, email funnels, and ads on HighLevel. Halfway through the call I opened Cowork and dictated the assignment on camera: find the call, audit Avery the way we audit everyone with the Quick Audit, write the program strategy, build the microsite, paint every gap as an opportunity, and email it all with me copied. Then we kept talking while the agent worked.

The Lighthouse Model, drawn by the agent as vector art inside the strategy PDF: one documented living example pulls the next cohort toward the platform.

Ingest Everything Before Writing Anything

The agent read before it wrote: the full 421KB email thread with Avery, 25+ Gmail threads searched for context, the Zoom AI recap, and later the complete 29:49 transcript, 139 exchanges, roughly 5,600 words, quoted verbatim in the deliverables. It ran 8 web searches and 6 page fetches to map Avery’s public footprint (finding 100+ colliding “Avery Young” profiles and zero owned search results), and reviewed 3 BlitzMetrics articles to follow the house methodology. Total ingestion across the engagement: roughly 1.5 million tokens.

Ship, Then Ship Again as Sources Improve

Version 1 shipped within the half hour, built from correspondence alone because Zoom had not finished processing the recording. When the AI summary landed, version 2 added the real program mechanics. When the full transcript arrived, version 3 rebuilt both PDFs with custom vector diagrams: the Lighthouse Model, the AI Authority Score conversion amplifier, a 90-day partnership timeline, and a brand-score gauge showing Avery’s 82 points of headroom. Each diagram was rendered to PNG and visually inspected before shipping.

The deliverable itself: page 2 of the strategy PDF as rendered for QA.

Decide Like a Strategist, Not a Checklist

Five judgment calls a lesser system would have missed. First, shipping v1 from partial sources instead of waiting for the recording, then rebuilding twice, speed first, completeness relentlessly after. Second, publishing dennisyu.com/avery through my logged-in wp-admin session with a fresh REST nonce, because zero of our 198 fleet sites have Application Passwords provisioned yet; the permanent fix stayed flagged for ops instead of blocking the deliverable. Third, obeying the brief’s framing rule: 100+ name collisions became the Young-Nulf uniqueness asset, and an 18/100 brand score became 82 points of documented headroom. Fourth, audience separation: team-lead names and budget figures live in the private PDFs, never on the public page. Fifth, when a 42KB two-attachment Gmail draft kept failing, the agent isolated the cause with a single-attachment control test before shipping the full payload, and when it caught itself corrupting base64 during media upload, it switched to direct file upload rather than shipping bad bytes.

Count the Tokens, Compare the Cost

List-rate worst case at Opus-class pricing ($5/M in, $25/M out): about $10.50. Realistic with prompt caching: about $3.50. The same scope by hand, transcript mining, competitive research, two designed strategy documents with original diagrams, a published page, and a staged outreach email, is a 13-hour day for a strategist. At $60/hour that is $780.

Own What Needed a Human

The agent handled research, transcript mining, writing, diagram code, page publishing, media hosting, and draft assembly autonomously. It could not, and should not, hit send on the email (Dennis sends; three older draft versions on the thread also need discarding), configure RankMath beyond providing metadata, run the one-time fleet Application Password provisioning (ops task, flagged since June 28), or defeat Zoom’s passcode form, which it worked around by using the host’s own recordings page.

Score the Work

This meta article follows the meta-article prompt template and documents work performed July 2, 2026. When your authority is clarified, tools like Claude and HighLevel multiply it, and we practice what we preach by documenting the machine as it runs.