Eating My Own Dog Food: How I Used the Framework to Write the Post About the Framework

Marc Millstone · article · 2026-05 source ↗ Reader 🖥️

AI Slop Cognitive Debt Human-in-the-Loop Mental Models & Understanding Writing as Thinking

The same risk-and-differentiation framework that tells engineering teams where to grant AI autonomy also governs knowledge work like writing: delegate heavily where errors are cheap and recoverable, retain full ownership where the work carries the argument's credibility, and use AI as an adversarial critic rather than a ghostwriter.

practical · reflexive case study / meta-demonstration — A practitioner walkthrough that applies the author's own AI-autonomy framework (from 'Don't Automate Your Moat') to the process of writing that same post. Functions as both proof-of-concept and practical guide for calibrating AI delegation in knowledge work.

Structure

Part	Role
Introduction: the meta-move	Frames the post as a self-referential test: the framework from 'Don't Automate Your Moat' must survive application to the writing of the post itself, or it is just rhetoric.
Full Automation: Citation mechanics	Demonstrates the low-risk, low-differentiation quadrant with footnote formatting -- cheap to fix, invisible to readers if wrong.
Collaborative Co-Creation: AI Gateway example and build-vs-buy framing	Shows the moderate-risk, moderate-differentiation quadrant where AI accelerates execution on design choices that the human has already validated.
Supervised Automation: Counterargument section	Illustrates higher-risk work where AI drafts but the human must verify every path -- the danger is not factual error but subtly unfair steelmanning that undermines trust.
Human-Led Craftsmanship: Opening, definitions, framework, evidence	The high-risk, high-differentiation quadrant: the opening narratives, axis definitions, framework genesis, and source verification all had to be owned outright because they carry the argument's credibility.
Cross-cutting practice 1: Using Claude as a critic	Shows that adversarial prompting (hostile persona + concrete criteria + permission to be brutal) surfaces structural and evidential weaknesses a solo writer would miss.
Cross-cutting practice 2: Voice discipline	Argues that AI's default register (em dashes, hedging, corporate-academic tone) must be actively edited out, and that voice fixes often reveal content weaknesses.
Takeaway	Synthesizes: delegate where error is cheapest, own where trust is at stake, and the friction encountered (drifting figures, misattributed citations, model voice) proves the framework rather than undermining it.

Key terms

quadrants
Four zones of AI autonomy defined by crossing business risk with competitive differentiation, each prescribing a different human-AI division of labor.
cognitive debt Cognitive Debt
The accumulated cost of not understanding AI-generated code or content you depend on -- analogous to technical debt but in the human's mental model.
build decision wearing a buy costume
The illusion that AI-generated code is 'built' by your team when it is functionally procured from a token-funded generation loop, inheriting the risks of a buy decision (opaque internals, dependency on vendor).
steelman
Presenting the strongest version of an opposing argument; used here as a quality bar for the counterargument section -- an unfair steelman is more dangerous than an obvious strawman.
tells AI Slop
Stylistic markers that reveal AI-generated prose: em dashes, rule-of-three lists, 'not just X, it's Y' constructions, words like delve/leverage/robust/nuanced.
mental model ownership Mental Models & Understanding
The ability to explain and defend every part of your work -- framework, evidence, citations -- without consulting the source, as an engineer would in an incident review.

Propositions

explicitThe risk-and-differentiation framework for AI autonomy applies not just to engineering but to knowledge work generally, including writing.conf: high
explicitThe parts of knowledge work that should be delegated most heavily to AI are those where being wrong is cheapest.conf: high
explicitAI's most valuable contribution to high-stakes knowledge work is adversarial critique, not content generation.conf: high
explicitcorroboratedAI-generated citations look clean but systematically drift in accuracy across drafts, requiring human verification against primary sources.conf: high
supports ←Inside a PhD Student's Research Workflow | Studying Parasocial Relationships with Alvina Lai · p12 — This document's implicit critical proposition (trusting peer-reviewed inputs doesn't guarantee Consensus's own AI-generated synthesis is drift-free) is supported by the established finding elsewhere that AI-generated citations 'look clean but systematically drift in accuracy across drafts, requiring human verification against primary sources.'
explicitAn unfair steelman is more damaging than an obvious strawman because it quietly undermines the entire argument's credibility.conf: high
explicitVoice editing and content editing are separable disciplines, but fixing AI voice often exposes content weaknesses because the generic register lets vague claims hide.conf: high
explicitcorroboratedA framework generated from a prompt reads like one; the seed idea must be human-originated to carry authentic weight.conf: medium
supports ←The best response to AI slop, infinite advice, and online noise is from Robin Williams · p8 — Franco/dog-food: 'a framework generated from a prompt reads like one; the seed idea must be human-originated to carry authentic weight.' Acunzo's 'a script the creator forgot to act — the words are there, but they're lifeless' is the same claim about the perceptibility of un-lived, non-human-seeded content.
implicitCollaborative co-creation works when the human drives design choices and the AI accelerates execution, not the reverse.conf: high
explicitThe friction encountered when using AI (drifting facts, misattributed sources, model voice) is not tool failure but evidence that human ownership is necessary.conf: high
explicitGeneric critique prompts produce useless feedback; effective AI critique requires a hostile persona, concrete evaluation criteria, and explicit permission to be harsh.conf: high

Arguments

The AI-autonomy framework is not just theoretical advice but a workable decision procedure for real knowledge work. strength: medium

Explicit premises:

The author applied his own four-quadrant framework to the writing of the post that introduces the framework.
Each quadrant prescribed a different human-AI ratio that matched the risk and differentiation profile of that piece of work.
The resulting post was stronger than either pure human or pure AI work would have produced.

Implicit premises:

A framework that survives self-application is more credible than one that is merely asserted.
Writing a persuasive argument is a valid proxy for the broader category of high-stakes knowledge work.

Inferences:

If the framework works reflexively on its own creation, it likely generalizes to other domains.

AI-generated content that looks correct can be systematically wrong in ways that only domain-expert verification catches. strength: high

Explicit premises:

The Knight Capital loss figure drifted across drafts because Claude pulled from whichever source appeared first.
A finding about senior developer maintenance burden got mapped to the wrong study.
A deskilling claim was phrased more broadly than the Lancet source supports.
Several sources were corrected, replaced, or cut after verification.

Implicit premises:

LLMs do not distinguish between primary and secondary sources or track which specific claim a source supports.
Clean formatting and confident tone mask citation-level errors.

Inferences:

Citation verification cannot be delegated to AI; it requires the human to hold a mental model of what each source actually proves.

Adversarial prompting is the highest-value use of AI in high-stakes writing. strength: high

Explicit premises:

Generic critique prompts produce generic, hedged feedback.
A hostile-persona prompt with concrete criteria surfaced real structural problems: conflated concepts, disputed causation in examples, overextended evidence claims.
Some pushback was accepted and improved the argument; some was rejected because it did not fit the specific context.

Implicit premises:

Solo writers with deadlines do not normally have access to skeptical reviewers on demand.
The value of critique is in having the notes at all, not in following every one.

Inferences:

AI-as-critic preserves human understanding (interrogative, not delegated) and scales the availability of adversarial review.

AI's default prose style actively harms persuasive writing by letting vague claims hide behind polished-sounding language. strength: high

Explicit premises:

Models default to a recognizable voice: em dashes, rule-of-three, 'not just X, it's Y', words like delve/leverage/robust.
Readers sense generic AI prose even when they cannot name the tells, and trust drops.
Rewriting for voice often forced content sharpening because the corporate-academic register conceals imprecision.

Implicit premises:

Trust is prerequisite to persuasion; a reader who detects inauthenticity discounts the argument.
Specificity and directness are both style choices and epistemic commitments.

Inferences:

Voice discipline is not cosmetic but structural: forcing plain, direct language forces the writer to know what they actually mean.

Narratives

The Knight Capital loss figure drifted across drafts because Claude pulled from SEC enforcement orders, Bloomberg, and other secondary sources interchangeably, citing different numbers on different passes.
→ Concrete evidence that AI citation handling is unreliable at the specific-claim level; demonstrates why source verification cannot be delegated.
Claude proposed the build-versus-buy framing; the author recognized its structural potential, and together they iterated until the phrase 'a buy decision wearing a build costume' emerged and the argument was reshaped around it.
→ Exemplifies genuine co-creation in the middle quadrant: AI contributes a conceptual seed, human judges its structural fit, and the final product is irreducible to either contributor alone.
The opening of the parent post -- an engineer who could not explain his own algorithm, a colleague paged about an undocumented service -- were scenes the author wrote from personal experience; a generated approximation would have read as inauthentic.
→ Anchors the human-led quadrant: authenticity in opening narratives is non-delegable because it is what earns the reader's initial trust.
Claude critiqued the post as a 'pro-AI, token-maxing CTO' and flagged that an early draft conflated cognitive debt as a risk problem with cognitive debt as a differentiation problem, weakening both.
→ Demonstrates the adversarial-critic prompt pattern and shows AI surfacing a real structural weakness the author had missed.
Claude told the author: 'This post is sound advice. It did not need sixteen footnotes to establish it.' -- but could not count to 18.
→ Wry anecdote illustrating that AI critique has genuine value even when the model makes its own errors; the point is having the notes, not treating them as infallible.

Questions addressed

✅ Does the risk-and-differentiation framework for AI autonomy actually work in practice?
Yes, the author demonstrates it working on the meta-task of writing the framework post itself, with each quadrant prescribing a different and appropriate human-AI ratio.
✅ Where does AI add the most value in high-stakes writing?
As an adversarial critic (stress-testing arguments, surfacing structural weaknesses) and as an execution accelerator in the co-creation quadrant -- not as a ghostwriter.
✅ What goes wrong when AI handles citations and evidence?
Figures drift, sources get misattributed to adjacent-but-wrong studies, and claims get stated more broadly than sources support. The output looks clean but is quietly broken.
✅ How do you prevent AI-written prose from undermining trust?
Separate voice editing from content editing; actively remove AI tells; recognize that forcing directness often exposes content vagueness.
⭕ Would the framework produce different quadrant assignments for a different kind of writing task (e.g., technical documentation, fiction)?
Not addressed. The post only demonstrates the framework on persuasive/argumentative writing.

Gaps

self-assessment biasThe author is grading his own framework by applying it to his own work. There is no external validation, no second writer replicating the quadrant assignments independently, and no reader-side evidence that the framework-guided post is actually stronger than a counterfactual version. The self-referential structure is rhetorically clever but epistemically circular.
Read next: Independent case studies applying the framework; A/B testing of framework-guided vs. unguided AI-assisted writing.
scopeThe demonstration covers exactly one genre (persuasive long-form essay) by one expert writer. No evidence is offered that the quadrant assignments would hold for other knowledge work types -- technical documentation, code review, strategic memos, creative writing -- where risk and differentiation profiles differ.
Read next: Apply the framework to diverse knowledge work outputs and check whether quadrant assignments remain stable.
mechanismThe post does not explain how a practitioner should determine which quadrant a given piece of work falls into. The author's assignments feel intuitive but the decision criteria are not made explicit enough for someone else to replicate them.
Read next: The parent post 'Don't Automate Your Moat' may provide the missing decision criteria; worth checking.
counterfactualNo comparison to what would have happened without the framework. The author asserts the framework guided his delegation decisions, but an experienced writer might have made similar choices by instinct. The framework's marginal value over general editorial judgment is not demonstrated.
Read next: Studies on structured decision-making vs. expert intuition in creative/knowledge work.
incompleteThe adversarial-critic prompt pattern is presented as highly effective but with only anecdotal evidence from one use case. The failure modes of this approach (model sycophancy despite hostile persona, critique that is adversarial in tone but generic in substance) are not explored.
Read next: Research on LLM sycophancy and the reliability of adversarial prompting for substantive critique.

Eating My Own Dog Food: How I Used the Framework to Write the Post About the Framework

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents