How I use AI

What this covers

Everything Mythic Works delivers: consulting and fix engagements, fractional CTO time, workshops and tutoring, websites I build, and the software I ship. It also covers what I publish under my own name.

What I use AI for

• Writing and reviewing code
• Drafting documents, plans, and assessments
• Research and background reading
• Analyzing logs, data, and codebases
• Drafting the writing I publish

That list is deliberately broad. Assume AI touched most of what I produce. The purpose of this page is not to claim otherwise. It is to tell you what that does and does not mean.

What it does not mean, and how the work is actually checked

AI is not the author and it is not the decision maker. You are hiring my judgment, and a model does not have any. My name on a deliverable means I am accountable for it. If it is wrong, that is on me, not on a tool.

I am not going to tell you I read every line of everything I ship. Nobody shipping at this pace does, and you should not believe anyone who says otherwise. What I run instead is layered review, so that no single pair of eyes is the only thing between a mistake and your production system. The layers are automated code review, tests, type checking, and security and license scanning. Anything touching authentication, payments, customer data, deletion, or a public interface, I read myself. The rule is proportional to blast radius, which is the same rule any competent engineering organization has always run on. No CTO has ever read every line their team shipped either. What they own is the system that catches problems, and so do I.

Generated code gets a stricter review than hand-written code, not a looser one. It arrives confident, plausible, and occasionally wrong in ways a person would not be wrong, so it does not get waved through for having been produced quickly.

I can always explain what a deliverable does and why it is built that way, without the tool in front of me. If I cannot, that is a defect, and I fix it before you see it.

Why these are my calls to make

I have been a professional designer, one way or another, for twenty-five years. That is the judgment you are hiring, and it is the thing a model does not have.

It also means I have worked under the rules that decide what data can go near a tool like this. COPPA, HIPAA, and GDPR are law, and they are the reason I ask what kind of data I am touching before I touch it rather than after. If your product handles children's data, protected health information, or EU personal data, the constraints are not the same, and neither is my handling. I have also been through SOC 2, which is an audit of controls rather than a law, and worth telling apart from the three above.

On quality

I do not ship slop.

That is easy to say, so here is what it means in practice. If part of a document or a module cannot justify why it is there, it comes out before you see it. Nothing goes out unverified, whether a person or a machine drafted it.

Volume is not the product. AI makes it trivial to produce forty pages where four would do, and a lot of what is being sold as AI-accelerated work is exactly that. Where these tools make me faster, you get the time back. You do not get more pages.

What I will not do

• I will not ship you anything that has not been through review appropriate to what it touches.
• I will not let a model make the judgment call you are paying me to make.
• I will not list an AI tool as an author or a contributor.
• I will not hand your material to a tool that has not committed against training on it. What that means in practice is in the next two sections.

The tools I use

Current tools, as of the date above. This is the whole list.

• Claude Code and the Anthropic API, for code, analysis, and drafting. This is where the work happens.
• The OpenAI API, for image generation only. No client material goes into it.

Neither vendor trains on the material I send them. On API billing that is contractual, and it is the default. On the Claude subscription it is an account setting, and mine is off.

I keep this list current rather than vague. If I adopt a new tool, I check its data-handling terms before any client work goes near it, and I update this page.

Your data

• It is never entered into a free tier, and never into a tool with training enabled.
• It is never used to train a model, by me or by a vendor.
• It is deleted from my active working copies at the end of an engagement on request. Encrypted backups age out on their normal retention schedule rather than being purged individually, and I will tell you what that schedule is if it matters to you.
• Both vendors hold inputs briefly for abuse monitoring under their own terms, separate from training. If your security team needs a zero-retention arrangement, both offer one and I will set it up for your engagement.
• You can impose stricter rules than these. Some clients require that no AI touch their work at all. That is a reasonable ask. Put it in writing before we start and it becomes a term of the engagement, at no extra cost and with no argument from me.

What you can ask for at any time

• Which tools were used substantially on your work
• The prompts and outputs behind a specific deliverable
• A walkthrough of any code or document I delivered, in enough depth to satisfy you that I understand it

You should not have to take my word for any of this. Ask.

Other people's work

I respect copyright, trademark, and trade secret, and I do not use these tools to get around any of them.

In practice: I check generated code for license contamination before it ships, the same as I check it for security. I do not prompt for imitations of a living artist's style or for anything carrying somebody else's mark. Your confidential material does not go into a tool, and neither does anyone else's. If I worked for your competitor last year, what I learned there stays there, and the tools do not change that.

What I publish

The writing on this site, my newsletter, and my posts are drafted with AI assistance. The idea is mine, the argument is mine, the edit is mine, and every factual claim is checked against a primary source before it goes out. Where a piece rests on a source, I link it. When I get something wrong, I correct it in place and say what changed.

Some cover and header images on my newsletter and posts are AI-generated, and I say so where it is not obvious. No AI-generated imagery goes into client deliverables. I do not use AI-generated video, anywhere, in anything.

The problems with AI are real

I am not going to sell you AI services and pretend the technology is harmless.

I wrote a long piece working through it section by section: the water and power that data centers pull out of the communities they land in, training sets built from work taken without consent, the labor conditions behind the cleanup, AI features wedged into products nobody asked to have them in, and what always-available chatbots do to people. It is here: The Problems with Data Centers and AI.

The through-line I landed on is not for or against. It is who decides and who pays. A few companies keep the upside of every one of those choices and hand the bill to whoever is nearest and least able to refuse. The fix keeps coming out the same shape: consent before they take, the people who profit carry the cost, rules for the powerful rather than homework for the powerless.

I still use these tools, including for that essay, and I check every number in what I produce against a source you can open yourself. The technology does not decide whether it is a machine for taking or a tool that helps a person think harder. The person using it decides. That is the whole of my position, and this page is me putting mine in writing so you can hold me to it.

Changes to this page

This policy is dated and versioned. Material changes are logged below. If a change affects an active engagement, I will tell you directly rather than expect you to re-read a web page.

Changelog

• 24 August 2026: First published.

Questions