The System Outgrew the Series

47 rules became 250+. The one file became 23. Three predictions broke. The boundary moved.

Three articles ago, I wrote that fact verification cannot be delegated to AI. Not to a better model. Not to a second pass. Not to a plugin. The human must click the link, check the number, read the original quote.

Then I built a system where AI cross-references every claim against curated sources and flags what's unverified. The same Claude. No model upgrade. The articles were snapshots of a workflow in motion. The brain/ system was built during the same three weeks as the series — each article's production teaching lessons that fed the next article's revision process. By the time Part 3 shipped, the system had already outgrown the claims Part 1 made.

Part 1 said I write my prose and Part 2 said the system has 47 rules. Part 3 said verification is irreducibly human. Each claim was true when I wrote it. Each one broke within weeks.

from 47 rules to 250+

The content-revision skill from Part 2 was one file with 47 rules covering voice, formatting, verification, and AI-sounding patterns. It worked. But as I shipped more articles, it wasn't enough.

The verification checklist grew. Prompting strategies needed templates. Evidence-gathering needed its own workflow. And the AI kept making mistakes that the skill file didn't cover because those mistakes happened in stages the skill file didn't reach: research, planning, distribution.

So the one file became a directory. brain/ now has 23 files:

• 5 skill files covering drafting, revision, plan creation, evidence correlation, and distribution. 250+ rules across the set.
• 5 prompt templates for each stage of the pipeline: drafting (personal and enterprise variants), editorial review, evidence correlation, distribution generation.
• 4 workflows that sequence the stages: research-to-publish, evidence-mapping, revision-cycle, brain-maintenance.
• 8 lesson retrospectives with commit-level analysis of real projects and revision passes.

I didn't design this structure upfront. Each file exists because a specific failure during article production demanded it. The evidence correlation skill exists because Part 3's fabrication taxonomy showed me what AI gets wrong. The distribution skill exists because early Twitter threads contradicted the blog posts they promoted. The editorial review prompt exists because the revision skill alone wasn't catching structural problems. The brain/ directory is the writing equivalent of a codebase's test suite and CI pipeline: structure that makes the output reliable without the human touching every line.

Heinrich argued that single skill files can't capture real domain depth. He proposed skill graphs: networks of small composable markdown files connected by links, where the agent navigates to exactly what the current task requires. In Part 2, I wrote that content-revision.md was "approaching that threshold" and that skill graphs were "the future." The future took weeks, not months. The brain is the skill graph, realized.

three predictions that broke

"I write my prose." Part 1 established the inversion: AI generates code, I generate prose. The first draft is human because the evidence is human. That was true when the evidence lived in my head. Once I built an evidence map from curated retrospectives and industry sources, the structured context was enough for AI to draft from. The missing ingredient wasn't a better model. It was structured input. Once I gave Claude the plan, the evidence map, the voice rules, and the retrospectives, it produced a structurally sound, factually grounded draft on the first pass. Not my voice yet, because the voice still needed editing, but the generation bottleneck moved from "AI can't write this" to "AI needs the right context to write this."

"47 rules." Part 2 described the content-revision skill as the system. 47 rules, each one a caught mistake. Same mechanism as CLAUDE.md. That mechanism didn't change. But the scope exploded. Drafting needed voice and structure rules. Evidence correlation needed cross-referencing rules against curated sources. Distribution needed platform-specific format rules for Twitter and LinkedIn. Plan creation needed its own evaluation framework. 47 rules in one file became 250+ across five. The mechanism scales.

"Verification cannot be delegated." Part 3 closed with the claim that fact verification is irreducibly human. The evidence correlation skill I built after that article reads the Obsidian vault (200+ digested industry articles), reads project retrospectives, cross-references every claim against those sources, and returns a status for each: VERIFIED, PARTIALLY VERIFIED, UNVERIFIED. It caught that a Crawshaw characterization was the retrospective's framing, not Crawshaw's exact words. It flagged metrics claims without primary source backing as UNVERIFIED. These are verification tasks. AI performed them against curated sources. Not full automation, but structured pre-verification that tells the human where to focus.

what moved the boundary

Not a model upgrade. The same Claude Opus that couldn't verify a URL in Part 3 now runs the evidence correlation pipeline. The difference is what sits around it.

Hashimoto's harness engineering from Part 2 is what happened at content-pipeline scale: every mistake, engineered away. 47 caught mistakes became 250+ encoded solutions across five skill files. Each one traces to a real failure during article production. The brain is what Anthropic's Skills Guide calls "execution design": the right skills loaded at each stage of the pipeline.

The boundary moved not because models got better, but because the harness got more sophisticated.

Boris Cherny ships 10-30 PRs a day running Opus at maximum effort. Always use the most capable model. But the model without the brain produced the fabrications documented in Part 3. The model with the brain produces drafts that pass evidence correlation on the first pass. Both matter: capable model and sophisticated harness. The operator's system matters more than the model's raw capability. A sophisticated harness on today's model beats a simple prompt on tomorrow's.

what's still true

The fabrication taxonomy from Part 3 holds. AI still fabricates metrics, distorts quotes, invents timelines, generates URLs that 404. The evidence correlation catches discrepancies against curated sources, but it can't browse live URLs, check live follower counts, or verify that a paraphrased quote preserves the original author's nuance.

This article proved it. The evidence correlation flagged a metrics claim as UNVERIFIED because I'd written a specific percentage with no primary source behind it. Category 6 from the taxonomy: metrics fabrication. The correlation caught it because the number appeared in no curated source. But it missed the file count: I wrote "24 files" when find brain/ -name "*.md" | wc -l returns 23. That's environmental verification, the kind the taxonomy says requires human work.

The human role shifted from "do all verification" to "verify the AI's verification." The correlation surfaces what it can check against curated sources. The human handles what it can't. The irreducible pass got smaller. It's still irreducible.

The 8 fabrication categories are model-agnostic and harness-agnostic. A better model fabricates less. A better harness catches more. Neither eliminates the taxonomy.

the meta-loop

This article was produced by the system it describes. AI read the retrospectives, built the evidence map from the Obsidian vault, drafted from the content plan, ran evidence correlation checking every claim against sources, ran editorial revision against the content-revision skill. The evidence correlation returned 22 verified claims, 4 partially verified, 4 unverified, and 2 inaccurate. The editorial revision caught anaphora patterns, a redundant blockquote, and a word count 161 words under the series minimum. I fixed the inaccuracies, verified what I could, rewrote the flagged patterns, and added the production detail you're reading now. The system did the structured checking. I did the judgment calls. Total production time from plan to revised draft: under 3 hours. The first three articles in the series each took 2-3 days of writing and revision. The brain didn't eliminate the work. It compressed it by encoding everything I learned from previous articles into the harness.

The same pattern as code. AI generates, human verifies. The inversion from Part 1 collapsed for structured content. When enough context is encoded in the harness, the code workflow and the writing workflow converge: AI drafts, human reviews. The difference between code and prose isn't the model's capability. It's how much structured context the harness provides.

Across all four pieces, the competitive advantage kept moving. Part 1: how fast you generate. Part 2: how fast you encode mistakes. Part 3: how fast you verify. Now: how fast your system learns to verify for you.

This is part 4 of a series on AI-assisted writing. Part 1: "AI Writes Code, I Write Prose" established the inversion. Part 2: "47 Rules, Each from a Real Mistake" built the system. Part 3: "Everything AI Fabricates When Editing an Article" mapped the taxonomy. Part 5: "I Gave My AI a Brain" builds the brain.

Built in Porto. Data from real articles, real revision passes, real fabrications caught — including the ones the brain caught this time.