Someone asked me a really good question yesterday: "Lynn, why are you against prompt augmentation... and don’t you do it in your systems?"
Great question. And the answer is… yes and no.
Let’s decompile this.
The problem with prompt augmentation is that it makes your prompting noisier. When you pack too much into the system prompt, you risk overloading the model with redundant or irrelevant information. That doesn’t make it smarter… it makes it dumber. (See: Cursor and the epidemic of stupid coding agents.)
You do need some augmentation in a tooled system, but even then, you're walking a fine line. The more you pollute the context, the less signal you have. You're not just using up space. You're chewing up reasoning cycles.
And no, the size of your context window doesn’t save you. It’s not about how much you can fit in—it's about what the model can meaningfully reason through. Let’s think about two critical but under-appreciated metrics – cognitive bandwidth and recursive depth.
LLaMA and Gemini both have large context windows. But LLaMA's recursive depth is shallow, because it flattens reasoning. Gemini has deeper recursion but heavier alignment constraints (including a few silly ones), which bottleneck its capacity for abstract reasoning.
Neither of them can process a prompt effectively if it's too noisy.
So: be selective. Know what your agent is for, and what outcome you're
steering toward.
Keep the roles legible
Augmenting user prompts is an even worse idea. Why?
Role confusion. You're tangling up thought processes, memory retrieval, tool selection, and user intent, all in the same needlessly messy block of text. It's unparseable.
If you could see what your context window actually looks like mid-session… you couldn’t do it, and you’re smarter than any LLM around.
Summarize this, indeed.
There’s no such thing as an AI hallucination. If it’s giving you an unexpected output, it’s because:
1️⃣ You gave it the wrong instructions
2️⃣ You were not clear enough in your directives
3️⃣ You pushed it off the map entirely
So how do you fix this?
1️⃣ Clear directives
2️⃣ Defined roles, rules, and objectives
3️⃣ Markdown everything
4️⃣ Your agents should be subjects, not objects. If you’re dealing with complex tasks, let them decide what data they need to consume, and when.
A good agent system is precise, traceable, and easy to reason through.
Do this right, and it works every time.