hawkes.aiKey ideas
- The one idea — What you put in the context — the right information, examples, and clear instructions — matters far more than clever phrasing or magic words.
- How it is shown — A "magic phrase" attempt failing, then the same task succeeding once the model is given the actual data and examples it was missing.
- The trap to avoid — Chasing prompt hacks and personas instead of supplying the missing information — superstition over engineering.
- What it sets up — You've engineered the INPUT — now control the OUTPUT: structure it.
Those viral 'secret prompt' lists are mostly superstition. What actually moves results: give the model the document, the examples, and a clear instruction. Engineering beats incantation.
The one idea
What you put in the context — the right information, examples, and clear instructions — matters far more than clever phrasing or magic words.
Two people, same model. One adds magic phrases — "world-class expert, take a deep breath" — and gets a mediocre result. The other adds the actual document, three examples, and a clear instruction — and gets exactly what they wanted. Same model. The difference wasn't the words. It was the context. So what actually works? Three things, all content, not phrasing. One: relevant information it can't know. Two: examples of what good output looks like. Three: clear, specific instructions. That's context engineering, and it beats every clever phrase — it fixes what the model was actually missing. Of the three, examples are the quiet superpower. Two or three examples of the format and style you want — few-shot — often beat paragraphs of description.
How it works — the demo
A "magic phrase" attempt failing, then the same task succeeding once the model is given the actual data and examples it was missing.
Here's an input, here's the ideal output: teaching by demonstration, which the model follows far more reliably than abstract description. Show, don't just tell. But magic phrases sometimes seem to work — why? Usually a crude proxy for real context. "Think step by step" nudges the model to show reasoning. "You're an expert" weakly narrows style. Pale shadows of giving the right instruction. When a trick works, better context works better. This makes prompting a debugging loop, not a spell hunt. Bad output? Ask what the model was missing. Lacked data? Add it. Wrong format?
The trap to avoid
Chasing prompt hacks and personas instead of supplying the missing information — superstition over engineering.
Why it matters — and what’s next
You've engineered the INPUT — now control the OUTPUT: structure it.
Show an example. Misunderstood? Sharpen the instruction. Each fix targets a real gap in the context, and the improvement is reliable — the opposite of trying phrases and hoping. The trap: collecting prompt hacks like charms — personas, secret phrases, threats, bribes — instead of supplying what's missing. It's superstition dressed as technique. The habit that compounds: ask, every time, what information or example the model needs — and provide it. Engineer the context; don't accumulate incantations. So context engineering — right information, examples, instructions — beats tricks, because it fixes real gaps instead of chanting at them. You can build a great input now. But there's a second half: to wire AI into software, you often need to control its output too — force it into a clean, machine-readable shape. Next: making it output clean JSON.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
Two people, same model. One adds magic phrases — "world-class expert, take a deep breath" — and gets a mediocre result. The other adds the actual document, three examples, and a clear instruction — and gets exactly what they wanted. Same model. The difference wasn't the words. It was the context.
So what actually works? Three things, all content, not phrasing. One: relevant information it can't know. Two: examples of what good output looks like. Three: clear, specific instructions. That's context engineering, and it beats every clever phrase — it fixes what the model was actually missing.
Of the three, examples are the quiet superpower. Two or three examples of the format and style you want — few-shot — often beat paragraphs of description. Here's an input, here's the ideal output: teaching by demonstration, which the model follows far more reliably than abstract description. Show, don't just tell.
But magic phrases sometimes seem to work — why? Usually a crude proxy for real context. "Think step by step" nudges the model to show reasoning. "You're an expert" weakly narrows style. Pale shadows of giving the right instruction. When a trick works, better context works better.
This makes prompting a debugging loop, not a spell hunt. Bad output? Ask what the model was missing. Lacked data? Add it. Wrong format? Show an example. Misunderstood? Sharpen the instruction. Each fix targets a real gap in the context, and the improvement is reliable — the opposite of trying phrases and hoping.
The trap: collecting prompt hacks like charms — personas, secret phrases, threats, bribes — instead of supplying what's missing. It's superstition dressed as technique. The habit that compounds: ask, every time, what information or example the model needs — and provide it. Engineer the context; don't accumulate incantations.
So context engineering — right information, examples, instructions — beats tricks, because it fixes real gaps instead of chanting at them. You can build a great input now. But there's a second half: to wire AI into software, you often need to control its output too — force it into a clean, machine-readable shape. Next: making it output clean JSON.