I Don't Fear AI Slop. I Fear Untold Stories with Zack Proser, WorkOS

If you care about someone or a team, you should encode your personal expertise and taste into an AI-powered 'factory' and give it to them, enabling them to c

Factories of Love and Grace: Encoding Personal Taste into AI-Mediated Creative Production

Abstract

This synthesis examines a practitioner account of encoding domain expertise and editorial taste into AI-mediated production systems, drawn from the work of an applied AI engineer at WorkOS. The central claim inverts the dominant anxiety about generative systems: the primary risk is not low-quality machine-generated output ("AI slop"), for which many mitigations exist, but the opportunity cost of stories that go untold because would-be authors do not identify as creators. Two systems are analyzed: Blog Bot, a Cloudflare-hosted, Slack-fronted pipeline that converts event transcripts into near-publishable drafts, and a four-hour personal application generating narrated bedtime stories for an audience of two children. From these, three transferable frameworks are extracted: asymmetric leverage, the spark/bellows/quench model of authorship, and a four-question quality rubric. Implications are discussed for human-in-the-loop design, authorship legitimacy, and reader trust.

1. Introduction

The deployment of large language models (LLMs) into creative and technical publishing workflows has produced a bifurcated discourse. One branch concerns output quality and the proliferation of undifferentiated machine text, commonly labeled "AI slop." The other concerns authorship, attribution, and the moral standing of AI-assisted work. The account synthesized here advances a third framing, which may be termed the factory thesis: an individual with accumulated expertise in a domain should encode that expertise into a tool and transfer it to people who do not ordinarily produce work of that type, thereby expanding who gets to create rather than merely accelerating existing creators.

The thesis rests on a specific biographical observation. The practitioner, a software developer and writer of fifteen and roughly two decades' standing respectively, describes himself as having been, prior to the availability of LLMs, effectively the only personal "content factory" he knew. He reports publishing an essay about his own writing process in 23 minutes following a demanding toddler bedtime routine, and notes that colleagues who asked how he wrote so quickly were given an answer rooted not in technique but in a persistent cognitive state: he was "always writing," continuously iterating on titles, framings, and sentences. The factory, in other words, already existed internally; it was simply not externalizable to anyone else.

This paper proceeds by establishing the intellectual context for the craft-versus-automation debate (§2), analyzing the two deployed systems and the authorship frameworks they motivated (§3), extracting implementation-level findings (§4), and discussing limitations and open questions (§5).

2. Background and Related Work

Two prior framings inform the analysis. The first is the assembly-line/craft metaphor contributed by Michael Grinich, founder of WorkOS, who compared Blog Bot to a factory line observed mass-producing cymbals. In that line, machines perform the forming and heavy labor while a craftsman listens to each finished cymbal, inspects it, and signs off before it receives a brand. Grinich's formulation - "craft runs through the entire line" - relocates craft from the act of fabrication to the act of judgment distributed across the process. The practitioner reports an initial aversion to this comparison, followed by agreement after reflecting on his own trajectory as a writer.

The second framing is adversarial and originates from a reader of the resulting output: increased publishing volume produced distrust, such that the reader reported flinching at each new artifact while attempting to determine whether it was machine- or human-authored. This response establishes the problem that the authorship frameworks discussed in §3.3 are designed to address, and it indicates that reader trust is a variable distinct from, and not automatically improved by, output quality or velocity.

3. Core Analysis

3.1 Blog Bot: Institutional Publishing Velocity

Blog Bot was constructed to raise publishing velocity across multiple teams at WorkOS performing innovative but under-documented work. The system runs on Cloudflare and explicitly encodes the builder's personal beliefs about what constitutes a good artifact - specifically, that it be engaging and pedagogically valuable. Colleagues interact with the system through a deliberately simple Slack interface, requesting an on-brand blog post derived from an event transcript. The pipeline auto-sources supporting context from Granola meeting recordings, event photography, and relevant Slack discussion threads, and includes a custom image-generation endpoint tuned to the design team's established visual style.

Outputs are characterized as 85-90% complete drafts, requiring human review, fact confirmation through built-in approval tools, and final editing before publication. The reported organizational effect was that publication rates rose across the company, and colleagues who had not previously shared their work began publishing regularly. This outcome is presented as evidence for the factory thesis: the limiting factor on institutional output was not available expertise but access to a usable production mechanism.

3.2 A Second Factory: Narrated Bedtime Stories

A contrasting case scales the factory concept down to an audience of two. Using a high-capability model variant, the practitioner built a personal application in four hours that analyzes photographs from his children's day, casts them as characters in a narrated story, and renders accompanying animation. The system includes real-time voice control, detecting a child's repeated utterance of "again again again" as an implicit replay request. Custom story generation from a spoken prompt - for example, a request to feature a specific character at a specific location - completes in approximately ninety seconds, and the practitioner reports an average of three replays per story. This case demonstrates that the factory thesis is not confined to institutional publishing contexts; the same pattern of encoding personal taste (in this instance, parental storytelling instinct) into a reusable system applies at the scale of a household.

3.3 The Spark/Bellows/Quench Framework

To resolve the tension between human and AI authorship exposed by the reader-trust problem in §2, the practitioner proposes a three-part model borrowed from blacksmithing. The spark is the conscious person with taste who determines, on the basis of lived experience of friction, pain, and joy, that a story is worth telling. The bellows is the continuous directive labor of shaping the work as it develops - deciding what to cut, what to expand, and what is incorrect. The quench is the final act of sealing and publishing, by which the author places their name behind every claim, link, specification, and data point in the work.

This model is paired with a four-question rubric for evaluating finished output: Is it true? Is it useful? Is it beautiful? Does it help anybody else? The practitioner further argues that if a reader cannot distinguish a fully hand-written piece from an AI-assisted one - even one produced over fourteen iterations across two and a half to three hours - there is, from the end user's perspective, no meaningful difference between the two. Authorship legitimacy is thereby relocated from the mechanics of production to the presence of an accountable human judgment at each of the three stages.

4. Technical Insights

Several implementation details generalize beyond the specific systems described. First, the asymmetric leverage pattern - a simple front-end interface concealing a non-trivial distributed backend - appears in both Blog Bot (Slack command triggering a multi-source Cloudflare pipeline) and the bedtime-story application (voice command triggering photo analysis, narrative generation, and animation rendering). This pattern allows non-technical users to invoke sophisticated capability without needing to understand or operate the underlying system.

Second, both systems preserve a human-in-the-loop checkpoint rather than fully automating publication: Blog Bot stops at an 85-90% draft requiring fact confirmation and manual approval, and the bedtime application's replay mechanism keeps a human (the child, interpreted by a parent-tuned system) continuously in control of iteration. This suggests a design principle: velocity gains are pursued up to the point of final judgment, not through it.

Third, context auto-sourcing - pulling from Granola transcripts, photographs, and chat history rather than requiring users to manually assemble inputs - appears to be a significant contributor to adoption, since it reduces the activation cost for colleagues who do not see themselves as writers. A limitation implicit in the account is that no quantitative benchmark is given for draft quality or editing time saved beyond the qualitative 85-90% estimate and the fourteen-iteration, 2.5-3 hour figure for a separate piece of writing.

5. Discussion

The factory thesis reframes AI-adoption strategy away from individual productivity gains and toward distribution of capability to non-producers. This has implications for how organizations might prioritize internal AI tooling investment: rather than optimizing tools for the fastest existing writer, the account suggests prioritizing tools that lower the threshold for colleagues who do not currently publish at all, since the marginal value of a previously untold story may exceed the marginal value of an existing writer's incremental speed.

The reader-distrust anecdote and the spark/bellows/quench response to it point to an unresolved tension in the broader field: increased production velocity can degrade trust even when per-unit quality is maintained, because the signal readers previously used to infer effort and credibility (visible production time, consistent authorial voice) is disrupted. The proposed rubric addresses this at the level of individual judgment but does not offer an organizational or technical mechanism for signaling to readers which checkpoints a human actually performed, leaving open a question for future work on provenance and disclosure design in AI-assisted publishing.

6. Conclusion

This account contributes a reframing of AI-adoption risk from quality control to opportunity cost, supported by two concrete implementations - an institutional publishing pipeline and a personal family application - that share an architecture of simple interfaces over complex, human-checkpointed backends. The spark/bellows/quench model and the accompanying four-question rubric offer a transferable vocabulary for maintaining authorship accountability under AI assistance. The practical takeaway for practitioners is to identify a domain of personal expertise or taste, encode it into a tool with a human approval checkpoint, and transfer that tool to people who do not currently produce in that domain, treating the resulting increase in output as evidence of previously latent creative capacity rather than as a dilution of craft.

Sources

• I Don't Fear AI Slop. I Fear Untold Stories - Zack Proser, WorkOS - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub