--- schema_version: 1 type: "essay" id: "essay-10-things-ai-has-taught-me" slug: "10-things-ai-has-taught-me" title: "Ten rules I use when working with AI models" summary: "Ten working rules for getting more from AI without confusing faster output with better thought." status: "published" visibility: "public" indexing: "index" categories: - "ai" - "thinking" - "building" tags: - "AI practice" - "judgment" - "verification" - "workflows" created: "2026-03-19" updated: "2026-07-29" published: "2026-07-23" canonical_path: "/writing/10-things-ai-has-taught-me" canonical_url: "https://rahilchamola.com/writing/10-things-ai-has-taught-me" preferred_citation: "Chamola, Rahil. “Ten rules I use when working with AI models.” Curated Randomness, 2026." reading_minutes: 10 authors: - id: "rahil-chamola" name: "Rahil Chamola" role: "author" legacy_slugs: - "10-rules-for-ai" - "10-rules-for-using-ai" series: id: "building-with-models" title: "Building with models" part: 2 related_slugs: - "co-writing-with-language-models" citations: [] --- # Ten rules I use when working with AI models My view of AI has become less dramatic the more I use it. The tools are impressive, but the quality of the result still depends on familiar questions: what am I trying to do, what information is relevant, how will I test the answer, and who is responsible for the decision that follows? I used to focus on the prompt and the response. I now pay more attention to the conditions around them. These are the ten rules that have survived that shift. ## 1. Define the work before choosing an AI model I start by describing the outcome, not the tool. What should the result help someone understand or do? What evidence may it use? Which constraints matter? What should happen if an AI model cannot complete the task honestly? Once those questions are answered, choosing an AI model becomes a practical decision about quality, speed, privacy, cost, and available tools rather than a referendum on which company is winning. This is also why I switch AI models only after comparing them on the work I actually do. A new AI model may be stronger overall and still add nothing to my workflow; another may appear less capable in a broad benchmark but perform better on the narrow task that matters. ## 2. Decide what context is relevant Context is the information available to an AI model while it works. Too little context produces generic answers, but too much creates a different problem: old claims, abandoned ideas, private material, and current instructions begin to arrive as one undated pile. I therefore try to provide the smallest complete set of information needed for the task. For an article, that may include the intended reader, my current position, the strongest objection, the evidence I trust, and the claims that remain uncertain. Selecting this material forces me to decide what is relevant before asking an AI model to reason with it, while also preventing personal history from becoming general-purpose fuel for every project. ## 3. Choose the AI model for the job I do not need one AI model to be the best writer, researcher, programmer, critic, and teacher at the same time. I need to understand which AI model performs each job well enough under the constraints I have. This has made me more comfortable with a method that works across AI models and is not tied to one AI model or provider. The brief, source material, review criteria, and final decision remain stable even when an AI model changes. That stability matters because AI model capabilities move quickly, while the purpose of the work usually changes much more slowly. ## 4. Separate making from judging Generating options and choosing among them are different forms of work. So are drafting an argument, checking its evidence, testing its strongest objection, and deciding whether it is ready to publish. When I ask one response to do all of these at once, an AI model tends to inherit its own assumptions and then approve the result they produced. I get better work by separating the stages. I may ask one pass for several possible structures, choose one myself, draft a bounded section, and then give the stable draft to a critic with a narrow question. The point is not to simulate a large team. It is to stop the first plausible answer from becoming the standard by which it is judged. ## 5. Test fluency and evidence separately A wrong answer can sound as calm and complete as a right one, which means tone cannot tell me whether a claim is reliable. The test has to match the work: a factual statement needs a source, code needs to run, a design needs to work in its actual interface, and a summary needs to survive comparison with the document it claims to represent. This also changes how I treat uncertainty. If a source is missing, the gap should remain visible. If the evidence supports only a narrow claim, the language should remain narrow. An AI model is useful partly because it can propose connections quickly, but speed makes it more important to distinguish what was observed from what has merely been inferred. ## 6. Use AI to exercise taste, not replace it AI can raise the minimum quality of a draft remarkably quickly. It can repair grammar, suggest structures, compare alternatives, and remove obvious repetition. None of that tells me which idea deserves to exist, which observation is unusually precise, or which technically competent paragraph has no reason to be there. Taste becomes more important when producing another acceptable version is cheap. I have to reject generic lines, preserve details that carry real experience, and notice when a smoother sentence has weakened the thought. An AI model gives me more material to choose from; the value still comes from the quality of the choices. ## 7. Save useful methods, not successful conversations When a session produces unusually good work, I do not assume that saving the entire conversation will reproduce it. The useful part is usually smaller: the brief, the order of operations, the examples, the checks, the rule for missing evidence, and the conditions the final result had to meet. Saving that method makes the lesson reusable without treating one long conversation as sacred. It also makes weak results easier to diagnose. I can ask whether the task was vague, the evidence was thin, the stages were badly ordered, or the test was too easy, rather than explaining every failure as a property of an AI model. ## 8. Use memory to record change, not freeze it Persistent memory sounds obviously useful until an old belief returns with the same confidence as a current one. A record that something happened is not evidence that it still defines me, and a position I once held should not become a standing instruction merely because the system retained it. Useful memory therefore needs time, source, scope, and a way to be corrected. I want to preserve decisions, reasons, unresolved questions, and the point from which work should resume, while leaving room for the decision itself to be reopened. Continuity is valuable only if it helps me notice how my thinking changed rather than quietly averaging those changes away. ## 9. Automate according to consequence, not capability I am comfortable giving AI broad freedom over work that is easy to inspect and reverse: cleaning notes, comparing options, preparing drafts, or running checks. I am more careful when an action affects another person, reveals private information, spends money, changes access, publishes a claim, or removes something that may be difficult to recover. The fact that an AI model can perform an action does not mean I have authorised it to do so. For consequential actions, I want to see the exact proposed change and make a clear decision. This is less convenient than full automation, but it keeps responsibility close to the point where the work leaves the draft and enters someone else's world. ## 10. Make the final judgment yourself Using AI to write articles helps me rapidly test my writing and ideas. I can turn one view into several structures, examine it from an opposing position, and locate places where the evidence does not support the confidence of the prose. However, the decision to choose one view over another rests with me. This is not because human judgment is mysterious or always superior. It is because an AI model cannot take responsibility for a claim, decide which trade-off I am willing to accept, or explain what I genuinely believe when the alternatives remain reasonable. If my name is attached to the work, I need to understand the argument and be willing to defend the final choice. AI has made more kinds of work possible for me and reduced the cost of testing an idea before I commit to it. The lesson has not been to keep the machine at a distance, but to become more precise about where its contribution ends and my decision begins. --- Creator: Rahil Chamola Canonical source: https://rahilchamola.com/writing/10-things-ai-has-taught-me Preferred citation: Chamola, Rahil. “Ten rules I use when working with AI models.” Curated Randomness, 2026.