AI is a tool. A named one, with limits.

What AI does here

• First drafts of articles, glossary entries, and tool copy. The model produces a draft from an outline I write. I rewrite every paragraph for voice, fact-check every claim against the sourcing policy, add the inline citations, and date the result. The byline is mine because the words and the responsibility are.
• Alt text and image descriptions. The model proposes a description from the file; I trim and edit. Faster than typing 1,300 alt strings; same accessibility outcome.
• Search and cross-link suggestions. The model surfaces related glossary terms and articles for the “keep going” rails on tool result pages. The connection has to read true to me before it ships.
• Spam triage on inbound mail. Light classification only — flag obvious bot submissions for me to review. No autoreplies.
• Translation seeds. The Spanish mirror starts as a machine draft, then a fluent native-Spanish reviewer rewrites for register and idiom. Nothing ships under /es/ that hasn’t had a human pass.
• Code, sometimes. Boilerplate, tests, and refactors. Every line is reviewed, run, and committed under my name. The generated code never includes data from a client engagement.

What AI never does here

• Train on your data. Tools run in your browser; the inputs never reach a server, let alone a training corpus. Studio engagements use API providers with the “no training” setting on; the policy is in the engagement letter, not the marketing copy.
• Sell prompts, embeddings, or transcripts. Nothing produced inside the studio — your invoices, your menu drafts, your meeting notes — gets resold, syndicated, or aggregated.
• Generate restaurant photos as real client work. The dishes in case studies are dishes a real photographer shot in a real kitchen. Hero shots are commissioned, not synthesized. Where AI imagery appears (illustrated diagrams, op-ed flat graphics), it’s labeled as illustration in the alt text.
• Write reviews, fake testimonials, or backdated “case studies.” Every review you read on this site has a person’s name and a verified context. Aggregate ratings only appear once five real ones exist.
• Make the call on a customer’s behalf. A model never sends a quote, signs off on a deliverable, or replies to your message as if it were me. The Window goes to my inbox; the answer is from me.
• Bypass the methods page. Any claim a draft surfaces gets a citation or it doesn’t ship. Halucinated statistics are how a library loses operators — it’s the failure mode I most try to avoid.

Tools and providers

The current stack, on the record, with versions on the date below:

• Drafting and editing. Claude (Anthropic), GPT-class models (OpenAI). API mode with training opt-out enabled; consumer-mode chats are used only for non-confidential generic drafts.
• Code assistance. Claude Code. Same training-opt-out posture. No client repository contents are pasted into a consumer-grade UI.
• Embeddings. OpenAI embeddings power glossary cross-links and search ranking. Embeddings are computed on public site content only.
• What’s never here. No autonomous “agent” runs over client data. No model-generated emails sent without my review. No generative-fill on photographs of real food.

Who reviews

Don Goldstein, the studio. Every page that shipped through an AI assist passed through my edit before publish. If I don’t agree with a sentence, it doesn’t make it to the page. The byline isn’t a model number.

If you spot something off

A claim that smells synthetic, a citation that doesn’t resolve, a tone shift that reads like a model rather than a person — don@muntin.digital or The Window with topic=ai. Reply within four business hours; corrections show in the changelog within two business days.

Sister surfaces

• Never — five things this studio will not do
• Security — nine claims about how data is handled
• Methods — sourcing and dating policy
• Changelog — what shipped, when
• Privacy — the legal version of the data policy