06-reference

alex lieberman content machine agentic workflow

2026-06-04·reference·source: X (long-form)·by Alex Lieberman (@businessbarista — founder Morning Brew / Tenex Labs / storyarb)
agentic-workflowcontent-machinemulti-agentcontent-opscritic-looplearning-loopsanity-checkpipeline

"I built a content machine" — Alex Lieberman (@businessbarista)

Why this is in the vault

A fully-specified, battle-tested (claims tens of millions in pipeline for Tenex) agentic content pipeline from a credible operator (Morning Brew founder). The founder flagged it as a process to follow + mine for workflow lessons — and it maps almost 1:1 onto RDCO's content engine (Sanity Check, /remix, /draft-review, /voice-match) and the pipeline-* multi-agent seats, with several specific patterns worth stealing. Converges with the same-day Google DS-STAR architecture share — both are verify/critic/refine multi-agent loops.

The machine (11 steps, "one pipeline, run end-to-end or step-by-step; a directory of skills")

Built by first mapping the traditional process (domain expertise first), then deciding where the human is needed (first mile = pick idea + provide context; final mile = fine-tooth review + sign-off) vs where AI runs (everything in between).

  1. The Oracle [AI] — mines Slack / Notion / call transcripts / Gmail for "spikes" (moments he naturally said something worth expanding) + an Internet Reader curates a selected external X/web feed. Ideas ≥6/10 written to "The Vault" (Notion idea DB).
  2. Select idea [Human]
  3. The Researcher [AI] — builds a sourced research-report.md (TL;DR, key facts w/ links, current developments, what's already been said, contrarian angles, open questions). Claims adversarially checked; fact separated from opinion.
  4. Interview Panel [AI + Human] — 6 world-class interviewer personas (Rogan, Stern, Barbaro…) ask 12-15 questions one at a time, each a different dimension, "never satisfied with vague answers — won't advance without 2-3 specific stories, real numbers, emotional specificity."
  5. Production [AI] — interview → raw .md (transcript, key stories, insights, quotables, emotional anchor, the "so what"). Raw file is sacred: exact words, never paraphrased away.
  6. Refinement [AI + Human] — reads custom style guide + past-feedback lessons + content-type spec, drafts in his voice from the raw file. #1 rule: "write like you're texting a friend."
  7. Writer's Council [AI] — 6 expert-reviewer personas (Shaan Puri, Morgan Housel, David Perell…) score the draft through their lens, splitting fixes into editorial (machine can rewrite) vs information-gaps (only the creator can answer → routed back to the interview panel).
  8. Revision Loop [AI] — iterate until Council scores 9/10.
  9. Repurposing Engine [AI] — 1 anchor → 10+ natively-formatted derivatives, each re-hooked per platform AND each held to the same full Council → 9/10 bar. "How two people produce like a hundred."
  10. Final revision [Human]
  11. Learning Loop [AI] — after approval, compares first draft vs final, extracts confirmed lessons → that creator's content-lessons.md. Every future first draft starts smarter; lessons override the style guide when they conflict.

Mapping against Ray Data Co

This is the strongest external blueprint the vault has for RDCO's content engine + the multi-agent pipeline pattern. Split into what RDCO already does vs the steal-able upgrades.

RDCO already does (validation):

Steal-able upgrades (the mined lessons):

  1. Writer's Council split: editorial-fix (auto-rewrite) vs information-gap (route back to the human/interview step). RDCO's critics (/draft-review, /design-critic, pipeline-critic) flag issues but don't cleanly separate "I can fix this" from "only the founder can answer this" and route the latter back. Highest-value add — it's the structural fix for critic loops that otherwise hallucinate-fill information gaps.
  2. Interview Panel as a tangible-extraction forcing function before drafting — interrogate the founder for 2-3 specific stories + real numbers + emotional specificity before writing. This is the mechanical enforcement of Sanity Check's tangible-not-abstract voice rule ([[feedback_x_voice_mismatch]]) — generate the concrete raw material first, don't let the draft go abstract.
  3. Hold derivatives to the same critic bar — /remix currently spins derivatives but doesn't run them back through the critic/9-10 gate. Lieberman does. Cheap quality lift.
  4. Personal-spike mining (the Oracle) — mine the founder's own channel history / call transcripts / Gmail for "moments worth expanding" → SC idea candidates. RDCO's /curiosity mines the vault; this adds mining the founder's own utterances.

Caveat: it's a content pipeline tuned to a media operator; RDCO's content surface (Sanity Check) is lower-volume, so don't over-build — cherry-pick #1 and #2, which are the load-bearing ideas.

Related