Levels of awareness in LLM use

Many people use AI, but very few use it well. I tried to organize the knowledge and tricks a user needs to acquire to move from naive use to expert use: five levels of increasing awareness, from L1 to L5, plus a starting level zero. Obviously these are generalizations and simplifications, not rigid boxes, and naturally you can be better and/or more aware in some aspects and not know others. And by the way, what really matters isn’t how much you spend monthly, but how much method you apply and how much experience you’ve accumulated.

L0, No knowledge

Uses ChatGPT, maybe without even creating a personal account; not aware that the same model has different versions with different performance. Doesn’t know how LLMs work, nor what their limits are, so asks questions without context and expects the truth, formatted according to their expectations on top of that. Two typical opposite outcomes: takes the LLM’s answer as gospel and is amazed by its dialectical ability, confusing form with content (epistemics: mistaking eloquence for knowledge, granting authority to whoever speaks well); or claims “AI is dumb” because it didn’t give them the answer they expected.

L1, The first real change

Starts preparing the prompt by providing context: indicates who the answer is for, provides examples if available, clearly specifies what they want to know. Quality rises considerably: this is the first step toward conscious use. Gets frustrated because, just as they start enjoying it, the free plan runs out of credits…

L2, Subscription and flagship models

Makes the big investment: the basic subscription, usually to ChatGPT or Claude.

Is aware that different versions of the same model exist, also because the first few times, when not paying attention, credit runs out quickly (the smarter a model version is, the more tokens—and money—it consumes).
If the question is difficult, uses the flagship model, the “thinking” or “adaptive” one, i.e. the model that thinks before giving an answer. Sometimes also does research independently or asks the model to do it. Quality rises again, quite a lot.
Usually has learned and internalized the sacred rule of AI use: “the LLM is a tool, don’t trust it, verify it” but if they haven’t learned it yet, the risk is perhaps even higher than at level 0.

L3, Savvy

This is the real leap in quality.

Knows how to calibrate effort depending on what’s at stake: quick questions for simple things, thinking model and research for important ones.
Creates small formulas to guide the model toward more grounded answers, meaning more anchored to verifiable sources and less to the model’s memory, so as to reduce hallucinations, such as “Short answer, explained simply, straight to the point, supported by source verification, zero sycophancy (a term similar to ‘compliance’ but more specific: the tendency of the model to agree with us, praise us, and not contradict us), don’t start taking action before we’ve clearly defined together what to do”, i.e. a set of instructions that partially curb some unproductive behaviors typical of LLMs.
If multiple sources are needed, knows that Deep Research mode exists, available in all major frontier LLMs, and that it introduces a major qualitative advancement because the research relies on a large amount of documentation drawn from the internet, even though it is costly in terms of time and tokens consumed.
If the problem is complex, senses that it needs to be broken down and analyzed in smaller parts, usually across different prompts, and also recognizes the limits of chats: has learned to resist the temptation to let them become endless.
Works even better on the prompt and context by asking the model itself to ask all the necessary questions, so as to avoid forgetting aspects important for the quality of the answer, or others considered obvious that are not at all.
Since by this point the LLM will have done research, reads directly the sources provided as links and verifies some of them, at least the most important ones, establishing a hierarchy based on origin (an important, recognized paper will carry more weight than a blog, for example).

L4, Master

Master level, for fundamental topics where maximum reliability and depth are required.

Usually has multiple subscriptions to frontier LLMs, so as to be able to choose the one considered best for a given task or a given phase of the research.
Builds Projects dedicated to the topic, because almost always the individual problem fits into a broader area of interest. The Project offers a specific memory and a repository of files that can be attached, accessible from all prompts within it; this way, integrates into the project the documents gathered independently, useful for tackling the problem. Makes use of the project’s specific memory: this reduces token usage for each new chat within it, because the LLM starts by reading that specific memory.
Checks whether specific skills exist to load into the LLM to make the model more specialized for a given task.
Once the answer or answers are obtained, cross-checks them with other frontier LLMs, which usually confirm the overall framework, but often bring out overlooked aspects or contradict secondary claims.

L5, Professional

There are many sophisticated techniques for improving the quality of LLM responses, including RAG systems, GraphRAG, databases, multi-agent tools, model fine-tuning, etc., but these fall outside the scope of this course. The field is vast and constantly evolving.

And you, what level are you at? In upcoming posts there will be deeper looks at what’s important to know and how to move up a level.