I've been on a tear lately trying to level up my understanding of agentic systems, schemas, vector stores, and the whole genetic-code-of-software thing that LLMs have cracked open. So I did what any reasonable human with a pulse does: I went down the rabbit hole. Reels, YouTube, blogs, Twitter threads from people who may or may not actually know what they're talking about. It got overwhelming fast, because everything is moving at warp speed and there are roughly forty-seven "right" ways to build the same thing.
So I put my product manager hat back on (the one thing I actually know how to do) and forced myself to solve fundamentals before chasing shiny objects. And along the way I figured out the single most useful piece of context I'd been leaving out of every prompt. It wasn't a schema. It wasn't a tool call pattern. It wasn't some clever RAG trick scraped (let's be honest, often stolen) from the open internet.
It was me.
Agents Are Only As Smart As the Context You Hand Them
Agentic systems can do genuinely beautiful things. Orchestrated chains of reasoning, sub-agents passing structured outputs around, results at light speed while consuming roughly half the planet's electricity (different post, different rant). But all of that breaks down if the context you feed in is garbage.
Here's the part nobody talks about: the model has read the entire internet. What it has not read is you. Your style, your gaps, your weird tolerance for ambiguity, the specific ways your brain skips steps. To the LLM, you are a black box wrapped in a prompt.
I am not a traditional coder. I can read code. I can mostly understand code. I cannot write more than about six lines without breaking something, because I'm not detail-oriented enough to spot the one misplaced bracket that a real engineer would find in twelve seconds. Give me a billion dollars and ten hours and I still probably wouldn't find it.
The Day I Stopped Pretending I Knew GitHub
For an embarrassingly long time, I prompted Claude like I was trying to impress it. I'd toss around words my engineering team used, nod along, and let it assume I knew what a rebase was. (Not speaking from personal experience or anything.) Turns out when you pretend to know GitHub, your deployments start quietly overwriting each other and you spend an evening wondering why your "fix" un-fixed yesterday's fix.
Or the classic: "Yes, go ahead and spin up a database." Cool. Now rebuild it three days later when you realize you needed a completely different shape of data because you didn't tell the model what you were actually doing.
The fix was almost stupidly simple. I started telling Claude, up front, that I don't know what I'm doing about 90% of the time, that I'm a product guy who can read code but not write it, that I'm prone to overthinking, and that it should push back on any instruction I give that doesn't make sense. Same outcomes. Way less pain. Way fewer 11pm "why is production broken" moments.
Tell the Model You're a Scatterbrained Lunatic
When people talk about context engineering, they almost always mean documents, schemas, tool definitions, memory layers. Fine. All useful. But the cheapest, highest-leverage context you can add is a blunt description of yourself. Your skill level. Your blind spots. The way you communicate when you're tired. The fact that you are (in other people's words, not mine, I think I'm pretty normal) a bit of a scatterbrained lunatic who is just thrilled he never has to write another status summary for a project that shouldn't have existed.
If you tell the model you're a senior engineer, it will talk to you like a senior engineer and skip the parts you actually need. If you tell it you're a product person who fakes the Git vocabulary, it will catch you before you nuke your branch.
And look, I get the privacy hand-wringing. "I don't want the AI to know that much about me." You use Google. You use a phone. That ship sailed, got scuttled, and was salvaged for parts a decade ago. At least here you get to choose what it knows and use it to actually make your life easier.
Give the model more of you. It works better when it stops having to guess.