How I Used AI to Make My Writing Better

I did not use AI to write these essays. The writing begins and ends with me: the memory, the argument, the joke, the frustration, the pattern, the lived example, the odd little connection that made the piece worth writing in the first place.

What I built was something different: an editorial system around the writing. I used AI while building parts of the tooling, mostly in JavaScript and Python, but the system is not a ghostwriter. It runs on my machine, not on this site. Nothing in it drafts a sentence for me, and nothing in it can reach into a published essay and change the prose. When it runs, it works more like an instrument panel than a writing engine. It helps me manage the full body of work, critique individual essays, detect recurring patterns, and find the places where a draft is too flat, too repetitive, too vague, too polished, or not quite itself yet.

The system has several parts:

There is a corpus management layer that tracks the entire essay library: titles, sections, categories, tags, publication status, word counts, summaries, pull quotes, card copy, metadata, and how each piece fits into the larger collection. That helps me see whether I am overbuilding one section, neglecting another, repeating a subject too often, or circling an idea that deserves its own essay.

There is an individual essay analysis engine that critiques a draft after I have written it. It looks at structure, pacing, title strength, opening strength, conclusion strength, clarity, rhythm, specificity, and whether the piece has enough lived detail. It does not replace judgment. It gives me another angle of attack.

There is a voice and consistency engine that keeps the work recognizably mine across a large body of writing: recurring motifs, tonal drift, favorite constructions, overused phrases, category fit, thematic overlap, and whether a piece still sounds like it belongs in the same house as the others. Bolted to it is a style telemetry layer that measures the things a writer can feel but not always see clearly: sentence length variance, paragraph length variance, lexical diversity, repeated vocabulary, passive voice, transition density, rhetorical flatness, and the places where the prose may be getting too smooth or too generic. The goal is not to standardize anything. The goal is to keep my very terrible habits from hardening into ruts while the essays stay human.

The tool also includes pattern detectors: word clouds, phrase frequency, semantic clusters, related-essay mapping, similarity checks, and topic overlap. Telemetry is where all of that becomes dashboards I can actually read, and Tests is the unglamorous companion that re-runs my editorial fixtures whenever I change how a measurement works, so the instrument panel does not quietly start lying to me. Sometimes the tool tells me a draft is repeating another piece. Sometimes it shows me that a throwaway line is actually the real essay. Sometimes it points out that I keep returning to a concept without naming it directly, which usually means there is more to dig out.

Then there is the Atlas Curioso, a curiosity and demand-mapping system built with AI-assisted code that ingested the full body of my work and connected it to the questions people are actually asking online. Priorities ranks what might deserve the next draft. Calibration is where I sit down and hand-label whether the machine's sense of related matches mine. Explorer lets me wander the neighborhoods by hand. Universe draws the whole library as terrain, so I can see which ideas have become towns and which are still one house on a dirt road. Fan-Out takes a question and breaks it into its logical branches, then marks which branches my essays already answer well, which ones they answer without ever surfacing it, and which are open territory. Yahoo mines real posted questions, questions only and never the answers, into a local pile I approve or reject one at a time.

All of that stays on my machine. When Fan-Out finds a genuine gap, the most it can do is start a draft: an empty noodling with a working title and the governing question sitting at the top, waiting for me. It cannot publish, it cannot link itself into other essays, and it cannot write the piece. AI helped me build the machinery, but the Atlas does not decide what matters or manufacture the answer. It shows me where real curiosity is accumulating so I can bring judgment, experience, language, and care to the part the machines keep missing.

The authoring side is deliberately dumb about AI. Admin is the cockpit: what is stale, what should probably run next, which long jobs are grinding away, and the button that tells search engines a page changed after I deploy. Essays, Pages, and Socials are ordinary editors where I type and press save, and this page was written in one of them. Spreadsheet handles the reference sheets and mapping profiles as tables, with a diff preview and a backup taken before anything is written. No model sits on the save path. Every word that reaches the site got there because I put it there.

Three smaller tools sit between the maps and the drafts. Joints shows which parts of the reference library a given essay actually touches, and which essays best illustrate a given entry. Related puts the public tag-based see-also list next to what the embeddings think is related; it is the only one of the three that can write anything, and the only thing it can write is that see-also list itself, which is metadata and never a paragraph. Draft check runs a work in progress through the same comparisons before it is anything at all, purely as analysis. None of them hands me a sentence to paste.

The last part is the one I care about most, which is the brakes. A save-quality gate watches what a revision does to a piece's measurable voice, and when a change looks like it is sanding the writing down toward competent nothing, it refuses the save until I read the reason and decide, on purpose, to override it. The semantic indexes and reports never leave my machine and are not part of what gets built and deployed. The live site calls no model at all; what it offers outward is a read-only view of published work for anything that wants to read it. And underneath all of it is the rule that made me build the thing this way: no AI ever touches live published prose. Drafts, metadata, and dashboards are the machine's business. The essays are mine.

The most useful part is that the tool helps me separate writing from editing. When I write, I can follow the idea. When I revise, the system helps me ask better questions. Is this section doing work? Did I explain the point but forget to make it felt? Did I use the same move three times? Is this line actually mine, or is it just competent filler wearing a nice jacket?

AI helped ease the creation of the tools that made the writing better, but not by becoming the writer. It helped me make something that became the critic, the indexer, the cartographer, the pattern spotter, the dashboard, the second set of eyes, and sometimes the little warning light that says: this part is technically fine, but it needs one more human thing.

The authorship is mine. The tools only helped me see the work more clearly, and not one of them has ever been allowed to write a word of what you are reading.