I’ve been to the artificial intelligence (AI) mountaintop… And I’ve seen the Promised Land! Mine eyes have seen the glory of the coming of the Borg…
In all honesty, I’ve learned so much about the capabilities and the liabilities of large language models (LLMs). As of today, there is only so much those models can achieve. When it comes to writing, I’ve learned that LLMs can’t replicate my voice in any way.
I’ve tried to look at the issue from so many angles. I’ve spent more than a month building an enormous writing codex. That codex is an amazing piece of vibe coding, if I’m being honest, but the output never comes close to being able to replicate my cadence, my prose, or my
I’ve tried Claude Code using Opus 4.8 Max, and I’ve also tried OpenAI’s Codex using ChatGPT 5.5 Max, but the results are the same, and by the same I mean that the output is always the same bland and obsequios slop; it’s almost impossible to force an LLM not to be milquetoast.
LLMs have constraints, which include terms like “context length” and “max output tokens”. The local LLMs like Google’s Gemma 4 12B and OpenAI’s GPT-OSS 20B are more constrained by these issues in comparison to cloud-based LLMs, but behemoths like Anthropic and OpenAI are also tied to these same limitations. There is no getting around the idea of “compute” issues.
I’m so tired of working with these LLMs. I’ve tried so many different iterations of LLMs and AI wrappers like Hermes and OpenClaw. I’ve wasted so much precious time. I need to reclaim it by becoming more human again. I thought that I was going to be a less attractive Seven of Nine clone, but that didn’t work. I’ve returned to writing as a salve to my robotic tendencies.
The following was written by ChatGPT 5.5 Max in my writing codex’s voice, about my writing codex, which is so meta…
The Writing Codex governs the pressure system behind my prose: sentence entry, sentence landing, punctuation, paragraph architecture, evidence, source notes, voice, form, and final review. The scope runs from raw drafting to public publication, across Markdown drafts, emails, replies, source-noted essays, and WordPress posts. The point is not to make every sentence sound intense; the point is to make every claim carry its actor, source, and consequence. The Codex protects the voice from platform polish, institutional altitude, and AI fluency that sounds clean because the hard parts were removed. A rule earns its place under pressure, and a rule that cannot survive the writing it governs deserves removal before enforcement.
The creation surfaces are the places where the writing enters the system before doctrine starts judging the result. Microsoft Word gives the blank page room to breathe, OneDrive keeps that raw surface backed up, Markdown becomes the source format for serious review, iA Writer becomes the reading and editing surface, and the WordPress block editor becomes the public publication surface. Finder keeps the local files visible, GitHub gives the project a recognizable version-control surface, and the repository keeps doctrine, evidence, governance, examples, and working instructions from turning into scattered memory.
The assistant layer and update machinery sit one level above the writing surfaces. ChatGPT, OpenAI Codex, and Claude Code help frame, retrieve, pressure-test, audit, and revise without becoming the author. AGENTS.md and CLAUDE.md define the agent contract; the AI launch packet, assistant brief, logic-debate protocol, retrieval map, and example bank translate that contract into working instructions. README.md, CODE_POLICY.md, DECISIONS.md, OPEN_QUESTIONS.md, and CHANGELOG.md keep the project’s governance, architecture, rulings, unresolved questions, and change record visible. Grammarly in Chrome belongs here as an advisory editor, not as a judge, with brand tones, writing preferences, and the Grammarly-based ledger within the Writing Codex recording where commercial grammar feedback helps the Codex and where the Codex rejects false authority.
The testing machinery begins only when a piece is ready for review. The draft stops expanding and starts answering for its claims. Swift, SwiftPM, Xcode, Apple Natural Language, Python, sentence-transformers, Bash, Pandoc, SQLite, plutil, AppleScript, Shortcuts, Pkl, SwiftLint, SwiftFormat, XcodeGen, Homebrew, npm, and OpenJDK support the local infrastructure. Vale, textlint, and yamllint form the blocking surface; Harper, LanguageTool, proselint, Ruff, and ShellCheck add advisory or code-static pressure. The source-fidelity checks — structure, fidelity, anchor-ledger, refinement-ledger, and source-note review — test whether the draft kept its shape and evidence. The claim and provenance checks — research-ledger, claim-ledger, and provenance-ledger — test outside facts, claim ownership, and suggestion history. The final-review checks — hostile-reader, failure-mode, LLM auditor packet and validation, token-landing, compile, compile-triage, and style-cases — test adversarial pressure, drift patterns, terminal weight, final repair order, and model behavior. The handoff and maintenance checks — partner-readiness, partner-smoke, session-index, session-packet, writing-session, lineage-watch, fixture-capture, session-closeout-smoke, real-session-fixture-smoke, chat-capture, voice-profile, longform-sources, longform-profile, longform-source-packet, workspace-surfaces, code-fidelity, code-static, code-cleanroom, toolchain, swift-smoke, and check-updates — test continuity, memory, corpus, workspace, code, native tooling, and maintenance. The tools surface pressure; judgment still belongs to the draft.
The overall project is the operating system around my authored work: an Apple-first, local-first command center on my iMac that lets drafting surfaces, assistant tools, source ledgers, final-review checks, and WordPress publication serve the same judgment. The project does not exist to automate taste, flatten judgment, or turn prose into a compliance checklist. The project exists to preserve witness, evidence, cadence, source trail, and ownership across the full path from first fragment to public page. AI may frame, retrieve, audit, pressure-test, and package, but authorship stays with the person whose name goes on the piece. The public standard is simple and severe: the piece can stand when its claims are locatable, its source trail is clean, and its cadence still belongs to my voice.
That took me about thirty minutes of prompting to finish it to my liking. Still, I think you can tell that it wasn’t written in my voice; it feels oddly antiseptic.
If you want to write efficiently and in good conscience, tools like ChatGPT, Claude, and Grammarly can help, but they can’t assimilate you as well as the glorious Borg.
Reply