--- title: "Write Medium-Quality Content" description: "Humans should decide what content says, AI should write most of the prose, and humans should review the result." url: "https://www.withgauge.com/blog/write-medium-quality-content/" author: "Ethan Finkel" published: "2026-08-12" --- # Write Medium-Quality Content Companies should write medium-quality content because it is the fastest way to explain every part of a product to AI systems and the people using them. By medium quality, I mean content where humans decide the subject, argument, length, and framing, then AI writes most of the prose. Humans review the draft, explain what does not work, and have AI revise it. AI writes the prose, but humans still control it. This division of work matters because companies now have to publish content that covers far more specific questions. People use ChatGPT and Claude to research products and answer questions. Developers use Claude Code and Codex to evaluate software, read documentation, and complete implementations. These systems retrieve information before the person gets an answer or the agent finishes its work. As AI handles more of this work, companies need enough content for it to retrieve an accurate explanation of what they do. Medium-quality content is the fastest way to create that surface area. ## AI retrieves narrow, specific information AI systems can turn a broad request into a series of narrow questions. Someone may ask which analytics product fits a healthcare company with a small data team. The model can search separately for HIPAA support, implementation time, warehouse integrations, pricing, and examples from similar companies before writing one answer. A developer can ask Claude Code or Codex to add analytics to an application without naming a vendor. The agent may retrieve installation guides, SDK references, framework compatibility, authentication requirements, and error documentation before deciding what to install. When [coding agents choose software](https://www.withgauge.com/blog/agent-led-growth/), focused product and documentation pages become part of the decision. In the example below, one agent task produced seven web searches about GitBook and Mintlify. The searches covered an OpenAPI reference, product analytics, migration complaints, monthly pricing, feature details, and a direct comparison. Each search asks for a different fact, and no general product page is likely to answer all of them. ![An AI agent making seven specific searches about GitBook and Mintlify documentation, pricing, analytics, and migration complaints](https://www.withgauge.com/assets/blog/write-medium-quality-content/agent-web-searches.png) A broad product page cannot answer every one of these questions well. It may mention that an integration exists, but it probably does not explain how the integration works for one framework, which versions it supports, or what changes when a team uses a particular data warehouse. Focused pages give AI a direct source for each answer. They also help people who search for the same narrow questions themselves. The more specific the question, the more useful a page becomes when it addresses that question directly. If a company's site does not contain the answer, the model has to infer it, rely on a third party, or leave the company out. Communicating clearly to AI therefore requires exhaustive coverage of the questions customers and their agents may ask. ## Exhaustive coverage is too expensive to write by hand The number of useful pages grows quickly. If a product has 20 important use cases and buyers ask 10 recurring questions about each one, the company has 200 combinations to explain. Some answers can share a page, but the full surface area still exceeds what most teams can write by hand. A writer can spend another hour improving the rhythm of one article, or use that hour to direct a second article that answers another customer question. Once the first article is accurate, clear, and useful, the second page usually adds more value than another round of prose edits. The quality curve explains why. Writing quality rises quickly when humans add a clear argument, sound evidence, useful examples, and a logical structure. After that point, each improvement takes more time and changes less about what the reader or AI system learns. ![Writing quality rises quickly before additional polish requires much more time](https://www.withgauge.com/assets/blog/write-medium-quality-content/quality-curve.png) Medium quality sits where the curve begins to flatten. The page says the right thing, answers the question, and gives the reader enough context to understand it. More work can improve the prose, but it will not materially improve the answer. Accuracy does not belong on this tradeoff curve. Medium-quality content still needs correct facts, defensible claims, and useful examples. The compromise applies to sentence-level polish, not to whether the page is true. ## Humans should decide what the page says AI can turn an outline into prose, but it cannot supply context it has never received. It does not know why a company built a feature, what customers misunderstand about it, which implementation detail matters, or which position the company wants to take. When AI receives only a title, it fills those gaps with the most common ideas in its training data. The result can sound competent while saying nothing new. The grammar is not the problem. The page lacks a reason to exist. Humans have the context that makes the page useful. They know the customer questions that keep appearing on sales calls, the constraints hidden in the codebase, the evidence behind a claim, and the exact point the company wants the reader to understand. Humans should do the hard work of deciding what the page says. They should choose the question, take a position, set the scope, provide the evidence, and explain how the argument should be framed. These decisions require firsthand context and original judgment. ## AI should write the prose, and humans should control it Once humans have supplied that direction, AI should do the hard work of turning it into complete prose. It can turn a detailed brief, transcript, or rough outline into a full article much faster than a person can write the same draft from a blank page. A useful input tells AI: - The question the page should answer - The reader who has that question - The claim the page should make - The mechanism that makes the claim true - The evidence, examples, or product details that support it - The claims the page should avoid - The target length and section structure With those inputs, AI can organize the argument, write transitions, expand examples, and produce complete sentences. That draft is not the final article. Humans still decide which sentences work, which arguments feel soft, which examples miss the point, and which sections need to be rewritten. Controlling the prose does not require rewriting every sentence. Humans can tell AI that an opening is too soft, a distinction is wrong, or a section does not follow the intended argument. AI can then rewrite the draft around that feedback. Humans control the result through direction and judgment, even when AI writes every sentence. On a technical page, reviewers may spend more time testing the commands than adjusting the introduction. On a comparison page, they may focus on whether each product is represented fairly. The depth of the review should follow the cost of being wrong. AI does not remove the writer from the process. It moves the writer's time away from producing sentences and toward deciding which ideas deserve to become pages, then shaping the drafts until they say what the writer intends. ## AI readers do not care who typed each word When AI is the main reader, it has no reason to value a sentence because someone wrote it by hand. It needs facts that match the question, answer it directly, and provide enough context to use the answer. A model gets the same information from a clear AI-written sentence as it does from a clear sentence written by hand. It may even find AI-written prose easier to consume when that prose states the answer more directly. This does not make quality irrelevant. AI systems can still misunderstand vague pages, repeat factual errors, and ignore content that never answers the question. The page still needs original substance, accurate details, and a clear answer. Humans do not need to choose every word themselves. For [pages built primarily for agents](https://www.withgauge.com/blog/serve-markdown-to-ai-agents/), using AI to write the prose is a rational choice. Documentation, implementation guides, compatibility pages, and narrow feature explanations exist to help a system complete a task. Clear and complete coverage matters more than a recognizable authorial voice. ## Some content should still be written more carefully Some content has a different job because the writing itself carries part of the value. Thought leadership, funding announcements, founder essays, social posts, and major product narratives communicate one person's ideas to a broad audience. In those pieces, the choice of words can reveal judgment. A carefully built narrative can change how the reader interprets the facts, and the writer's voice is part of what the audience came to read. Those pieces deserve more direct attention from the person whose name appears on them. Long-tail content has a more specific job. A page about one feature, integration, implementation detail, or narrow buyer question succeeds when the reader finds a correct and complete answer. Spending several additional hours on its prose usually produces less value than covering the next unanswered question. Humans should spend more time writing and editing when the voice is part of the argument. For long-tail content, this process makes the full detail of a product available to people and AI systems that are searching for it. ## A practical medium-quality workflow The process can stay simple: 1. Humans choose one specific question that deserves its own page. 2. They dictate or outline the answer, including the claim, context, evidence, and boundaries. 3. AI turns that direction into a complete draft. 4. Humans check the facts, reasoning, positioning, examples, and prose. 5. They explain what does not work, and AI revises the draft. 6. Humans approve the result, then move to the next useful question. This workflow keeps human effort on the parts that require judgment. It also creates a natural quality floor because the draft begins with real context instead of a title and a request to fill space. ## This article is medium-quality content This entire post took less than an hour to create. Ethan, the writer, used dictation to tell these ideas to his computer. AI turned that direction into a complete article, then Ethan reviewed the draft and explained what he wanted changed. Ethan did not write 1,500 words himself. He controlled the prose by saying that the opening was too soft, the argument had the wrong hierarchy, and the distinction between writing and control was incorrect. AI rewrote the article after each round of feedback. The transcript below contains the original dictation that supplied the argument. Later task instructions and editorial feedback are omitted. > We're going to write a blog post called "Write Medium-Quality Content." This blog post is going to make the core argument that the ideal quality of content to write is medium-quality content. That means content that is written with human instruction in terms of what to write, the specific lengths to take, and the framing of the argument, with AI executing the actual prose and verbiage of the post. > > We also would recommend that a human review, edit, and shape the post after the AI writes it, but it's not necessary that a human writes every single word. The key rationale for this is that being exhaustive and covering all angles is extremely important. We see that AI searches for very niche, hyper-specific things. When you're trying to communicate messages to AI, the more hyper-specific you are and the more focused you are, the more effective your writing is, or the more likely your writing is to be consumed. > > Now, it's way too cumbersome and time-consuming to write a bunch of hyper-specific pieces by hand without any AI assistance. It just is too much of a body of work to take on. If you just had AI write all of it without any human input, it wouldn't really answer the question quite the right way because it doesn't have the relevant context on what argument it should be making. Since that's a real-world, novel idea that you're trying to portray, our recommendation is to have the AI do the hard work of creating the prose for the argument, for the content, and have the human do the hard work of coming up with what the content should be writing about. > > For example, the piece I'm currently dictating to you is a great idea, a great example of medium-quality content. I am doing the hard work of dictating the piece and coming up with what to do. The AI, in this case, is doing the hard work of writing the entire article to cover this lens or that, to actually write the prose that is published. > > Another reason why medium-quality content works well is that if your audience is AI, AI's not really going to know the difference between what's written by humans and what's written by AI. In fact, it may even prefer some of the prose that AI generates instead of the prose that I happen to think is prettier. With that kind of constraint of AI being the consumer of content, it's totally rational for AI to be the writer of the content. > > Now, if your main audience is humans and you want a human to consume this content (for example, a piece of thought leadership or something you're going to be blasting on socials, a funding announcement, or anything of that ilk), those are great things for people to write. You want sharp, personified ideas that are coming from one person's mind to many people. If you're writing something that's like a long-tail documentation page that's going to detail a specific feature based off of the constraints that are in your codebase, AI is probably going to be consuming it anyways. It makes sense for you to have AI write it and not care too, too much about the prose being your exact prose. Granted, you definitely want to sharpen it up. You want to make sure that it matches what you want to publish, but the importance of you cherry-picking every individual word is not very high. > > This is the article we're going to be writing and publishing on the blog. The sentiment that I went through here in our dictation is the key sentiment we'll be covering. It's actually kind of meta that the way we're writing this article, or advocating for medium-quality content, is by having us do the process recommended for generating medium-quality content.