Alex Lieberman’s Claude Content Machine: A Better System for AI Content Without the Slop

AI Content Does Not Need Another Prompt. It Needs an Editorial System.

Our first reaction to Alex Lieberman’s Claude “Content Machine” was that it gets the diagnosis right. Most AI content is not bad because the model cannot write a grammatical sentence. It is bad because the company gives the model nothing worth saying, then mistakes speed for a content strategy. The result is familiar: interchangeable LinkedIn posts, agreeable newsletters with no edge, search articles that summarize what everyone already knows, and a growing pile of content that does not create authority or distribution.

Lieberman built Morning Brew into a business newsletter with more than four million subscribers before its majority acquisition by Business Insider in 2020. That background matters because audience businesses learn an expensive lesson early: a publishing cadence alone is not an asset. The asset is a reliable system for turning insight into attention, attention into a direct audience, and a direct audience into repeatable distribution. Claude can accelerate parts of that system. It cannot substitute for it.

[INFO_TABLE] Product/Service: Alex Lieberman’s Claude Content Machine Primary platform: Claude Core job: Turn internal and external signals into voice-consistent content drafts Idea source: “Oracle” scans internal data and the web for idea spikes Quality control: Interview panel, voice/style files, editorial council, author approval Supporting tools: Notion, Slack, Claude Code, YouTube, X, LinkedIn Price: TBA – workflow operating cost depends on model usage, tooling, and human editorial time [/INFO_TABLE]

The important detail in that table is not Claude. Claude is the engine, not the operating model. The machine works because it separates four jobs that companies routinely mash into one vague instruction: finding a signal, developing a real point of view, writing in a recognisable voice, and deciding whether the piece should be published. Once those jobs are separated, AI becomes easier to manage. It has clear inputs, clear handoffs, and a clear point where a person owns the final call.

The Real Product Is an Idea Supply Chain

Content teams often describe their constraint as “we need more content.” That is almost never the actual constraint. The bottleneck is that they do not have a dependable way to identify ideas with commercial relevance before drafting begins. They begin with a blank document, a generic topic, or a keyword exported from a tool. Then they wonder why AI produces generic output. A model cannot create a differentiated argument from an undifferentiated brief.

Lieberman’s system starts upstream with an Oracle that scans internal data and the web to generate idea “spikes.” The language is useful. A spike is not a finished content brief and it is not a mandate to publish. It is a potentially valuable signal: a repeated customer question, a useful internal observation, a shift in how a market talks about a problem, a pattern emerging in audience conversations, or a tension worth investigating. This is the work that makes content a discovery function rather than a formatting function.

For an operator, the implication is straightforward. Your company already produces raw material that competitors cannot copy: sales-call friction, customer support themes, onboarding failures, product usage patterns, internal debates, founder observations, community conversations, and lessons from shipping. Most of it disappears into Slack, meeting transcripts, CRM notes, and the heads of people doing the work. The first job of an AI content workflow is to retrieve that material before it is lost.

Step One: Let the Oracle Find Signals, Not Finished Opinions

There is a temptation to treat AI monitoring as a content vending machine. Feed it the web, ask for trending topics, and publish whatever comes back. That is the fast route to becoming another company reacting to the same public conversations as everyone else. An Oracle is more valuable when it identifies patterns for a human to interrogate. It should surface the raw tension, evidence, and source material that could become an opinion-not write the opinion before the company has earned one.

This distinction protects distribution. Platforms reward novelty inconsistently, but audiences remember useful specificity. A strong post about a pattern buried in your customer conversations can travel because it carries insight people have not already seen. A polished take on a broad news event may receive polite engagement, but it rarely builds a reason to follow the author. Visibility compounds when the market starts to associate a person or company with a particular lens.

Step Two: Use an Interview Panel to Extract the Real Idea

The moment this workflow clicks is the interview panel. Instead of asking Claude to expand an idea spike into a post, Lieberman uses a structured interview process to pull the underlying experience, examples, counterarguments, and point of view out of the author. That is where the value sits. The AI is not being asked to impersonate a founder’s judgment from a handful of adjectives. It is being used to ask better follow-up questions and preserve the material generated in the conversation.

Across the companies we study, this is where most “AI slop” can be prevented. Slop is often not a writing failure. It is an input failure. The author has not supplied an unusual observation, a concrete story, a hard-earned trade-off, or a position that excludes some alternatives. AI then fills the empty space with the statistically safest version of the topic. It sounds coherent because it is coherent. It sounds generic because it has nothing specific to carry.

An interview panel changes the unit of production. The unit is no longer a prompt. It is a captured argument. The system should force details into the record: what happened, what surprised the operator, what changed their view, what the common advice gets wrong, what evidence supports the claim, and what a reader should do differently. Once those answers exist, drafting is the easy part. More importantly, the draft has a chance of containing something worth distributing.

• Capture the original signal rather than a cleaned-up summary of it.
• Record the author’s actual interpretation and the evidence behind it.
• Identify the tension or contrarian element that gives the idea an edge.
• Extract examples, language, and operating details that only the company can provide.
• Decide what the reader should understand or change before generating the draft.

Step Three: Treat Voice Files as Operating Assets

Voice and style files are sometimes framed as a cosmetic layer: preferred words, banned phrases, sentence length, and a few examples of past work. That is necessary but incomplete. A useful voice file should capture the author’s intellectual posture as well. What do they notice? What do they distrust? How do they frame trade-offs? What kinds of claims require proof? What would they never say because it is too vague, too promotional, or simply not true?

This is why a brand voice guide and an author voice guide should not be the same document. The company needs consistent standards. But the founder, executive, or employee advocate needs an identifiable point of view. If every person in the company publishes through one flattened brand template, the content may look consistent while feeling anonymous. That is a poor trade. Authority is an asset precisely because people attach it to a credible source with a distinct perspective.

We would build these files from real artifacts: strong past posts, memorable customer emails, podcast transcripts, internal memos, sales-call explanations, and writing that produced a useful reaction from the market. Wispr Flow and similar voice-input tools can make it easier to capture thinking before it is polished away, especially for operators who articulate ideas more naturally in conversation than in a blank document. The goal is not to automate personality. The goal is to preserve the raw material that makes personality visible.

Step Four: Make the Editorial Council a Gate, Not a Decoration

The editorial council is the part many teams will skip because it feels slower. That would be a mistake. Generative AI lowers the cost of drafting so sharply that the new constraint becomes judgment. A council gives the workflow a structured review layer before publication: Does the piece make a real claim? Is the argument specific enough? Does it reflect the author’s perspective? Is there evidence? Does it belong on this distribution channel? Has the model introduced a clean but unearned sentence that needs to be cut?

That does not require turning every LinkedIn post into a committee meeting. It means encoding editorial standards before the volume rises. A founder might remain the final approver for flagship posts. A content lead may approve newsletter drafts. Subject-matter leaders can validate technical claims. The critical principle is that approval is intentional and accountable. Automation can prepare, route, compare, and draft. It should not quietly publish unsupported claims because nobody had time to read the output.

The Feedback Loop Turns Output Into a System

Lieberman’s workflow includes a feedback loop because the first version of a voice system is rarely accurate enough. Every edit is useful training material. When an author removes a phrase, adds a sharper example, changes the order of an argument, or rejects an idea as off-brand, the workflow should preserve that decision. Over time, the system can become less dependent on repetitive correction because the standards are documented rather than trapped in one person’s head.

This is the leverage most teams miss. They use AI to generate more drafts, then manually repair the same mistakes every week. That is not a system; it is a faster treadmill. A system records what failed, updates the instructions, improves the source material, and makes the next production cycle stronger. The useful metric is not how many pieces AI generated. It is how much repeatable editorial judgment the company has successfully encoded without removing human responsibility.

That feedback loop also makes the workflow resilient when distribution changes. Search is changing under AI Overviews. LinkedIn formats change. X changes. Newsletter inboxes remain crowded. A company that relies on a single channel is exposed whenever the channel rewrites the rules. A company that captures its ideas, voice, audience questions, and editorial decisions in owned systems can adapt the packaging while keeping the underlying intellectual asset.

The Same Pattern Works Beyond Founder Content

The architecture behind this Content Machine shows up in other useful AI workflows. One creator pipeline scans Slack community channels, extracts questions with AI, archives them in Notion, clusters similar issues, drafts social posts, and routes them through Buffer. The author retains final approval. That is a clean model for employee advocacy or community-led marketing because it converts real audience demand into content without surrendering control of the company’s voice.

Another system applies the same logic to content gap analysis. It combines competitor data with Search Console and Google Analytics, clusters opportunities according to business impact, and outputs page-level refresh recommendations. The value is not a massive keyword-gap spreadsheet. Nobody needs another spreadsheet with thousands of theoretically possible topics. The value is a prioritised decision: which pages matter, why they matter, what should change, and how the work connects to a business outcome.

A Claude Code pipeline takes the process further by connecting Claude to Ahrefs MCP data, using roughly 23 custom skill files, and saving artifacts at each stage for auditing. It targets publish-ready SEO drafts in six to twelve minutes while maintaining human review. The timing is less interesting than the trail of artifacts. A production system is easier to trust when the team can inspect the source data, the intermediate decisions, the generated draft, and the edits that made it publishable.

These are different implementations of the same operating principle: AI should turn messy inputs into clearer actions while a human retains ownership of judgment. Slack questions become a content backlog. Search and competitor data become prioritised refreshes. A founder’s experience becomes a distinct argument. The common gain is not “more content.” It is a better conversion rate from existing company knowledge into discoverable, distributable assets.

Where the Content Machine Still Needs Discipline

There is no magic in calling a workflow an Oracle or an editorial council. Bad inputs still produce bad ideas. Weak competitor selection can distort a content-gap analysis. Internal data can be incomplete. Community questions can overrepresent the loudest users rather than the most valuable opportunities. A voice file can become stale. And an approval layer can become performative if reviewers rubber-stamp drafts because the volume is too high.

There is also a real tension between speed and accountability. The more automated the pipeline becomes, the easier it is to create a flood of plausible work that nobody has truly reviewed. That is why saved artifacts, explicit approval stages, and clear ownership matter. They create an audit trail and make it possible to locate where a weak claim entered the system. This is especially important for companies using AI to support technical, financial, or high-trust categories where a confident-sounding error can damage authority quickly.

We would resist the urge to measure this workflow by volume alone. More drafts do not automatically create more reach. More scheduled posts do not automatically build authority. The real test is whether the company is becoming better at finding differentiated signals, articulating them in a recognisable voice, and placing them across channels without losing quality. Distribution beats content, but only when the content carries a reason for the market to pay attention.

How We Would Build This Inside a Growing Company

Start smaller than the full machine. First, choose one high-value distribution surface: a founder’s LinkedIn account, a weekly customer newsletter, a technical blog, or a community content stream. Second, create one reliable signal source, such as sales-call notes, Slack questions, product feedback, or customer interviews. Third, run structured interviews against the strongest signals until there is a library of genuine arguments. Only then should the company invest in detailed voice files, an editorial council, automation, and cross-channel adaptation.

• Signal layer: collect recurring questions, internal observations, customer language, and market changes.
• Idea layer: use AI to cluster patterns and identify spikes worth developing.
• Insight layer: interview the operator or subject-matter expert until the piece has a defensible point of view.
• Voice layer: apply documented examples, preferences, exclusions, and argument patterns.
• Editorial layer: review for evidence, specificity, channel fit, and approval.
• Learning layer: save edits and rejected outputs so the workflow improves rather than merely repeats.

The outcome is not a robot author. It is a company with better editorial memory. Ideas no longer disappear after a meeting. Strong language does not remain trapped in a sales call. The founder does not have to begin every post from zero. And the content team stops treating every asset as a one-off production task. That is how visibility compounds: not through endless publishing, but through a system that repeatedly turns proprietary experience into work people can discover, remember, and trust.