{"slug": "ai-policy-html", "title": "ai-policy.html", "summary": "A developer has published a detailed AI usage policy explaining how they use large language models for grammar checking and code implementation but not for generating ideas or writing drafts. The policy includes a JavaScript library that uses jsdiff to show readers exactly what the author wrote versus what adjustments an LLM made. The author acknowledges that LLM-influenced phrasing patterns like \"Not X but Y\" and the rule of three have become more common in their recent writing, raising questions about where inspiration ends and generation begins.", "body_md": "To help with understanding the usage of AI, I have a small JavaScript library (which you're welcome to take) that leans on jsdiff to show you what I wrote and what adjustments an LLM made in this policy.\n\n# How I use AI for writing prose\n\nI use LLMs. I do not use them to generate ideas. I do not use them to write my drafts. I'm happy to use them as researchers, they're strong search partners, but I don't trust their interpretations of research so I read all my sources.\n\nI do ask it to check spelling, grammar when I've reached what I believe is my final version; it most often notes where a sentence I've revised a few times is actually missing words or some other defect of that nature.\n\nThere's one article where you will see an LLM was definitely used for phrasing, \"AI Doesn't Lighten the Burden of Mastery\". I was, in that article, trying hard to get out the message that you're going to need to work hard and understand detail, even if using an LLM. The first paragraph got restructured by an LLM more heavily than I can justify calling a grammar change; AI didn't write a word of that paragraph, but it's fair to say it wrote the structure.\n\nThat idea of an LLM generating a structure? That's where the trickiness is for me. For example, there are two prose patterns that LLMs love that seem to land well with people: Not X but Y, and following the rule of 3s. I've written more prose with these phrasings in the last 12 months than in the previous 12 years, because it seems to land well with audiences. Where does this land on the LLM-inspired vs generated argument? I don't know.\n\n# How I use AI for Code and Client Work\n\nIn code and client work, I'm comfortable using LLMs to speed up the implementation, though I strongly believe in human design, human conversation. I do honestly think that asking an LLM to knock out some Terraform I can describe is a better deail for my clients, and it's faster for me.\n\nI think of it like using a power tool to cut and shape wood: there are some human-worked details that power tools lose, like chisel marks on the surface, but most modern people don't think this disqualifies a master carpenter from being called a master. It means the carpenter's balancing purity and pragmatism.\n\n## How I use LLMs in my Personal Software\n\nA project like [Andrew](https://github.com/playtechnique/andrew) does get touches from LLMs; I\ndon't mind at all that it can help me get details of a finite state machine correct. I don't ask for\nideas. I write these projects to learn things, so I do not lean on LLMs much; knowledge is won through\nwork.\n\nProjects that are AI-heavy are clearly labelled as such. I'm experimenting and learning the right way to use the tools, and I recognise that learning these tools requires using them.\n\nTo help with understanding the usage of AI, I have a small javascript library (which you're welcome to steal) that leans on jsdiff to show you what I wrote and what adjustments an LLM made. You can see it on this policy.\n\n# How I use AI for writing prose\n\nI use an LLM. I do not use them to generate ideas. I do not use them to write my drafts. I'm happy to use them as researchers, they're strong search partners, but I don't trust their interpretations of research so I read all my sources.\n\nI do ask it to check spelling, grammar when I've reached what I believe is my final version; it most often notes where a sentence I've revised a few times is actually missing words or some other defect of that nature.\n\nThis paragraph I've written primarily for the diff view, an LLM did not suggest it. I feel free to disagree with the editing suggestion. For example, in the above change of 'LLMs' to 'LLM', the LLM actually suggested leaving LLMs plural and instead changing 'it' to 'them' in the following paragraph.\n\nThere's one article where you will see an LLM was definitely used for phrasing, \"AI Doesn't Lighten the Burden of Mastery\". I was, in that article, trying hard to get out the message that you're going to need to work hard and understand detail, even if using an LLM. The first paragraph got restructured by an LLM more heavily than I can justify calling a grammar change; AI didn't write a word of that paragraph, but it's fair to say it wrote the structure.\n\nThat idea of an LLM generating a structure? That's where the trickiness is for me. For example, there are two prose patterns that LLMs love that seem to land well with people: Not X but Y, and following the rule of 3s. I've written more prose with these phrasings in the last 12 months than in the previous 12 years, because it seems to land well with audiences. Where does this land on the LLM-inspired vs generated argument? I don't know.\n\n# How I use AI for Code and Client Work\n\nIn code and client work, I'm comfortable using LLMs to speed up the implementation, though I strongly believe in human design, human conversation. I do honestly think that asking an LLM to knock out some Terraform I can describe is a better deail for my clients, and it's faster for me.\n\nI think of it like using a power tool to cut and shape wood: there are some human-worked details that power tools lose, like chisel marks on the surface, but most modern people don't think this disqualifies a master carpenter from being called a master. It means the carpenter's balancing purity and pragmatism.\n\n## How I use LLMs in my Personal Software\n\nA project like [Andrew](https://github.com/playtechnique/andrew) does get touches from LLMs; I\ndon't mind at all that it can help me get details of a finite state machine correct. I don't ask for\nideas. I write these projects to learn things, so I do not lean on LLMs; knowledge is won through work.\n\nProjects that are AI-heavy are clearly labelled as such. I'm experimenting and learning the right way to use the tools, and I recognise that learning these tools requires using them.", "url": "https://wpnews.pro/news/ai-policy-html", "canonical_source": "https://playtechnique.io/ai-policy.html", "published_at": "2026-05-25 18:58:20+00:00", "updated_at": "2026-05-25 19:39:23.575082+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "generative-ai", "ai-tools", "ai-ethics"], "entities": ["jsdiff", "AI Doesn't Lighten the Burden of Mastery"], "alternates": {"html": "https://wpnews.pro/news/ai-policy-html", "markdown": "https://wpnews.pro/news/ai-policy-html.md", "text": "https://wpnews.pro/news/ai-policy-html.txt", "jsonld": "https://wpnews.pro/news/ai-policy-html.jsonld"}}