Updated - 7 August, 2026
I publish architectures for working with AI, and I sell some of them. It would be strange to do that without saying plainly how this work is made, what I earn from, and where my interests and my arguments point in the same direction. This page is that account. It is maintained rather than published once; the date above is the last revision.
This is not a response to a feature. The standard it describes is the one I already apply to the architectures I publish, where a system is required to declare what it executed against what it merely narrated, and what it enforces against what it only honours. It would be a strange thing to hold software to that and not myself.
How I earn income
Published capabilities. ORGANISMIC publishes governed architectures — folders of plain text you load into an AI, read, run, and keep. One is free and permanent. The rest are sold at a fixed price with a perpetual licence. There are no subscriptions, and there is no tier of my work that expires if you stop paying.
Books. Two are in progress and will be sold. Their opening movements are published free.
Installation work. I am developing an engagement practice for firms that want the same property installed at organisational scale. As of this writing it has no clients, and I will say so on this page until it does.
Nothing else. No advertising, no sponsored posts, no affiliate links, no paid placements. If that changes, it will be disclosed here first and labelled at the point of use.
The conflict, stated plainly
These essays argue for a position that my products occupy. I argue that capability you own is worth more than capability you rent, and I sell owned capability. I argue that firms must become legible before AI can be leveraged at their core, and I am building a practice that does exactly that work.
I do not think this invalidates the arguments — I built the products because I believed the arguments, not the reverse, and the essays predate most of what is for sale. But you should know it while reading, and I would rather tell you than have you discover it.
Two commitments follow. I will not make a claim in an essay that I could not defend if I had nothing to sell. And where an essay's argument leads to something I sell, I will say so in the essay rather than at the end of a funnel.
Use of AI tools
I use frontier models extensively and specifically. Here is how.
Thesis and argument. Not delegated. Every essay descends from a claim I hold and can be held to. I have been argued out of positions by a model and into others — twice, that changed a published piece — but the decision about what to claim is mine, and a model that produces a persuasive paragraph in support of something I should not be claiming is producing exactly the failure I am watching for.
Drafting. Collaborative, and substantially so. I work with Claude the way my own writing recommends working with a model: I hold the claim, the material, and the standard; it holds the horsepower. It drafts, it argues back, it proposes structures I had not considered, and it is regularly overruled. Some published passages were machine-drafted at my direction and then judged and ratified by me. I could not reliably separate them from the hand-written ones after the fact, which is why I think a percentage is a poor instrument for this question.
Revision and pressure-testing. I give a model the finished thing and instruct it to attack the argument, find the unsupported claim, and name the objection a hostile reader would raise first. Several essays carry a concession or a hedge that exists because that pass found something.
Verification. Every factual claim goes through a pass whose standing instruction is to treat the manuscript as hostile. It has removed statistics I liked, corrected a claim I had already published, and forced hedges into sentences that read better without them. Where a claim cannot be sourced, it is narrowed to what can be supported or marked as unsupported.
Structural and editorial work. Formatting, consistency checks across a corpus, catching a term used two ways in two documents, converting tables that will not render. Unglamorous and constant.
Images and diagrams. Some illustrative plates and diagrams on organismic.org are AI-generated and are identified as such where they appear. Diagrams that make factual claims about an architecture are drawn from the architecture's own registry rather than generated, because a generated diagram can be beautifully wrong.
What I do not use it for. I do not use a model to decide what is true. I do not publish a claim because a model phrased it well. I do not use it to produce volume for its own sake.
What I put into models
My own writing, my own drafts, my own unpublished work — routinely. The benefit outweighs my concern, and I would rather say that than pretend to a caution I do not practise.
Client and third-party confidential material — never, unless it is already public. Where I work with a firm I will follow their AI policy, use their internal platform if they have one, and ask rather than assume.
Personal data belonging to other people — never.
Verification and corrections
Mistakes get made. When I find one, the correction is stated rather than quietly edited, and where a published claim was wrong the change is recorded in the piece's revision log with the reason.
I have already done this once in public: an essay carried a claim about prior art that turned out to be false when I read the source in full, and the claim was removed and the removal logged. I expect to do it again.
If you find an error, tell me. I would rather be corrected than right.
Sponsorship, partnership, and attribution
I take no sponsorship and run no advertising. If that changes, it will be labelled at the point of use and disclosed here.
Where my work derives from someone else's framework or talk, I say so by name and I claim only what I added. Where a partnership exists, it will be stated on the artifact itself. I will not imply a relationship that does not exist, including by voicing an asset as though a creator had endorsed it.
Privacy and data
I collect email addresses to deliver what you asked for and to tell you when it is updated. Nothing is sold, rented, or shared. Analytics are anonymised. If you want your address deleted, ask and it is done.
Human collaboration
This is a one-person firm and most of the work is mine. Where that changes — an editor, a fact-checker, a collaborator on a specific piece — it will be named here and on the work.
Elvin Garcia · ORGANISMIC · organismic.org