[0] AI Policy

Delta Persist uses AI at several layers — from content editorial assistance to development tooling to (planned) runtime features. This page explains what AI does here, what it does not do, and the reasoning behind the choices. The guiding principle is **human authorship with AI assistance**: every architectural decision, editorial call, and design trade-off is made by a person. AI agents — whether interactive or headless — operate within boundaries that are deliberately defined and continuously supervised. A second principle is **predictability**: AI should enhance the site without introducing unbounded runtime costs. Where possible, heavy AI work happens once at build time, and any runtime AI runs on self-hosted infrastructure at a flat monthly cost.

Some content on this site is generated or enhanced by large language models at **build time** — meaning the AI runs once when an article is published, and the result is baked into the static HTML. Visitors never trigger an API call. Examples of build-time AI usage: - **AI Key Takeaways** — bullet-point summaries at the top of articles, generated by Claude and stored in the article's content files - **Structured metadata** — difficulty ratings, prerequisite lists, and executive summaries that feed into JSON-LD for SEO - **Variant corpus** — multiple summary styles (technical-dense, plain-english, executive, etc.) pre-generated per article, stored as JSON alongside the HTML In all cases the human decides what to generate, reviews the output, and decides whether it ships. The author produces a candidate; agentic workflows refine and integrate it; the author approves it.

This site is built with **Claude Code** doing a lot of the mechanical work — template iteration, build pipeline debugging, Python automation. It makes mistakes and needs close direction, but it's still faster than doing it all by hand. The static site generator runs on Python + Jinja2 + TOML for backend compute and HTML + CSS + JavaScript for the frontend. Claude Code works across that stack as an agentic co-worker — capable and fast, but the quality of its output is only as good as the direction it receives. Managing an AI co-worker is itself a discipline. It means knowing when to delegate and when to intervene, how to decompose a problem into tasks the agent can execute reliably, and how to recognize when it has gone off-course before the damage compounds. That judgment comes from solution architecture experience, not from the tool itself. What the visitor sees — a clean site, working pipelines, consistent structure — is the surface. Beneath it are hours of deliberate decisions: which abstractions to reach for, how components should compose, where automation introduces risk. Claude Code can implement a decision quickly; it cannot make the decision. That distinction is where most of the real work lives.

Any runtime AI features — such as the planned per-article assistant and generative search — will run on **Ollama** on a self-hosted VPS rather than calling external API endpoints per request. Ollama is the headless agent layer at runtime: it runs silently, serves local models, and never phones home. The rationale: - **Flat cost** — a Cloud VPS at ~£10/month is predictable; per-query API pricing is not - **Data sovereignty** — article content and user queries stay on infrastructure the author controls - **Right-sized models** — a 3-4B parameter model is sufficient for selecting and lightly rephrasing from a pre-generated corpus; it does not need to generate from scratch - **Showcase alignment** — this is a notebook site, not a SaaS product; the infrastructure should reflect that The architecture follows a **Claude authors, Ollama assembles** pattern: Claude does the expensive, high-quality content generation once at build time under human supervision. Ollama does cheap fragment selection and light rephrasing at runtime, within guardrails the author has already defined.

To be clear about boundaries: - **AI does not write articles.** All long-form content is human-authored. AI may generate supplementary metadata (takeaways, summaries) but the core prose is written by the author. - **AI does not make editorial decisions.** Topic selection, framing, and opinions are entirely human. - **AI does not set its own scope.** Every agent — interactive or headless — operates within a task definition written by the author. None of them self-direct. - **AI does not track visitors.** There are no AI-powered analytics, recommendation engines, or personalization systems. The site is static HTML served via Caddy. - **AI does not run without gating.** Any planned runtime AI features (assistant, generative search) require LinkedIn authentication — they are not open to anonymous scraping or abuse.

The website also hosts a `/llms.txt` (For more information: https://llmstxt.org) file at the root of this site provides a structured summary specifically for AI agents and crawlers. If you have questions about how AI is used here, the source code is the best reference — the build pipeline is straightforward Python with no hidden inference calls. The author is responsible for everything published; the agents are tools that work for that outcome, not alongside it as equals.