Tired of people not reading your docs? The problem isn't your writing. The problem is your toolchain doesn't care about quality — and neither does anyone else's. Here's one that does.
Tokens are a dime a dozen. You can generate a thousand pages of docs before lunch — and the market is drowning in the output. What's changing isn't the tools — it's what people do with the result. The question that used to matter was can AI do this? The question that's starting to matter is did a human care?
For the better part of two years, teams pointed AI at their docs and walked away. The output was technically complete — covered the topic, answered the question, said nothing. Not a technology failure. An abdication failure. The human decided what to generate, pressed enter, and left the judgment seat. The machine kept running.
The damage isn't the slop itself. Mediocre docs have always existed. The damage is what the flood did to the reader.
Points AI at a topic, presses enter, ships the output. Technically complete, semantically empty. The human decided what to generate but not what to say.
Learned to pattern-match the mediocre. Instinctively, the way you stop reading a publication after it burns you three times. Tokens are cheap. Trust is not.
The reader's distrust changed the economics. Generating docs got cheaper every quarter. Getting someone to read them got harder. The bar didn't fall — it rose.
Pirsig's thesis: Quality emerges when the person doing the work is present in it. The mechanic who cares produces a different result. Not a different wrench — a different relationship to the work.
The mechanic can hand the wrench to a machine that turns it perfectly and indifferently. Caring didn't get easier. It got harder. The machine removed the excuse of difficulty — and with it, the forcing function.
So how do you keep a human in the seat when the machine can do the work without one?
You don't solve this with willpower. "Use AI responsibly" is the quality equivalent of "eat less and exercise more" — true, useless, and immediately forgotten under pressure. You solve it with architecture.
Drafting, formatting, rendering, publishing. Instant, tireless, indifferent.
What to say, what to cut, what to leave out, what earns the reader's time.
You can't publish without an editorial decision. Not policy. Architecture.
That's the claim. Now watch it work.
You're reading the proof.
This page was built by four MCP servers — context, flowbot, xhtml-tools, and wiki. They handle everything that isn't taste: files, workflow enforcement, rendering, publishing. The human's only job is judgment. That division isn't maintained by discipline — it's enforced by the system.
The system locks the LLM in a read-only cage during the thesis phase. No writing, no publishing — only reading and compressing. Inside this cage, each iteration follows the same loop:
Each iteration repeats until the human is satisfied. Three produced the thesis:
Human → LLM: "the thesis is 'raise the quality ceiling'"
Human → LLM: "back off — say 'here's how to not generate slop'"
Human → LLM: "tired of people not reading your docs?"
Will sent feedback via Slack: "the message isn't clear... what is this age of quality? what threshold did we cross?" The LLM can't read Slack. This feedback entered through the human — who decided what mattered — and reshaped the thesis before the LLM had any say. His question drove the shift from the abstract (raise the quality ceiling) to the concrete (tired of people not reading your docs?).
When the human shifted the flow out of read-only — unlocking the LLM to draft, render, and publish — the page failed four times before it worked. Each failure is a publish-screenshot-evaluate cycle where the visual result didn't meet the bar. The LLM assumed a color scheme existed — it didn't. A custom scheme registered in one server but was invisible to another — MCP servers don't share memory. Dark backgrounds rendered with invisible text because the system assumed dark mode while the reader was in light mode.
Four failures to get to a working page. Each followed the same iteration loop — the LLM published, the human evaluated, diagnosed, and directed the next attempt. The system made trying again free.
Margarida read the published page. "What does an iteration mean? What does a failure mean? The whole HOW THIS PAGE WAS BUILT section is confusing — you should define your terms." The vocabulary the builder took for granted wasn't obvious to the reader. Her feedback entered the same loop: the human evaluated it, directed the rewrite, and the section you just read was rewritten to define its model before demonstrating it.
Thesis + outline + voice + human editorial decisions
Themed Confluence page, dark surfaces, readable
Four failures, each diagnosed, each retried, none silent
The quality lives at the boundary between human judgment and machine production. The failures prove someone was standing there.
Same toolchain. Same four servers. Same capability. If the human had typed "write me an essay about AI quality and publish it":
No Will's feedback. No three iterations. No earned direction.
No tense rules. No Pirsig. No structural reasoning.
One page_scaffold call. No screenshot. No retry. Default scheme.
Technically complete. Covers the topic. Nobody reads it.
The difference between this page and slop is whether someone was standing at the boundary.
Resolves where things live. Every file path, every read, every write goes through context. The LLM never touches the filesystem directly.
Enforces cognitive order: thesis before outline, outline before draft. Narrows available tools at each phase. The cage that prevents premature action.
Turns markdown into themed XHTML. Color schemes, card layouts, dark surfaces. The human never writes markup.
Pushes content to the site, takes screenshots for QA. The final boundary between draft and reader.
4 of 19 servers. 270+ tools. See the full system.
This page was built with the system described above. AI handled production. The judgment seat was shared — the author held the system, Will reshaped the thesis, Margarida clarified the vocabulary. The quality — or lack of it — is the result.