hawkes.aiKey ideas
- The one idea — A skill is a packaged, reusable capability — instructions, and often a tool or reference file — that an agent loads on demand when it's relevant, instead of cramming every capability into context up front.
- How it is shown — An agent with a shelf of skill-cartridges, pulling in only the "format a spreadsheet" skill when the task needs it and leaving the rest unloaded.
- The trap to avoid — Thinking a skill upgrades the model, or that you should load them all at once — a skill is packaged context/tools, and its whole value is loading only when relevant.
- What it sets up — Skills add capability on demand — but how do you inject control into the loop itself?
How do you give an agent a hundred capabilities without stuffing them all into its context? You package each one — and it loads only what the moment needs.
The one idea
A skill is a packaged, reusable capability — instructions, and often a tool or reference file — that an agent loads on demand when it's relevant, instead of cramming every capability into context up front.
Last group ended on a promise: the harness isn't fixed — you can extend it. The first and simplest way is skills. A skill is a reusable, packaged capability an agent loads only when it needs it. Let's see what that means. A skill bundles instructions — how to do a task well — with, often, a tool or a reference file, under a name. Think of it as a capability cartridge: when a task calls for it, the agent loads it into context. Nothing new in the model — just packaged, reusable know-how. The key phrase is on demand. You don't pour every skill into the window at once — that bloats the context and rots it. Instead, the agent sees a short list of what's available and loads a skill's full contents only when relevant.
How it works — the demo
An agent with a shelf of skill-cartridges, pulling in only the "format a spreadsheet" skill when the task needs it and leaving the rest unloaded.
Progressive disclosure: a lean window, capability a reach away. This solves a real tension. You want an agent that can do many things, but can't fit all those instructions in the window at once. Skills let capability scale without the context paying for it up front. Add a hundred; the agent still loads only the one or two a task needs. And notice what a skill actually is: mostly text — instructions the model reads — plus maybe a tool it already knows how to call. It's the prompt-is-the-program idea, packaged and reusable. Write it once, and any agent with the harness can load it. Capability becomes shareable. The trap comes in two forms.
The trap to avoid
Thinking a skill upgrades the model, or that you should load them all at once — a skill is packaged context/tools, and its whole value is loading only when relevant.
Why it matters — and what’s next
Skills add capability on demand — but how do you inject control into the loop itself?
One: thinking a skill upgrades the model — it doesn't; it's packaged context and tools. Two: loading them all to be safe — which puts you back to a bloated, rotting window, the problem skills exist to solve. The whole value is loading only what's relevant. So a skill is a reusable capability the agent loads on demand — instructions and tools, packaged, pulled in only when relevant. That extends what the harness can do. But skills add capability; they don't control the loop. How do you inject your own logic into the cycle? Next: hooks.
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
Last group ended on a promise: the harness isn't fixed — you can extend it. The first and simplest way is skills. A skill is a reusable, packaged capability an agent loads only when it needs it. Let's see what that means.
A skill bundles instructions — how to do a task well — with, often, a tool or a reference file, under a name. Think of it as a capability cartridge: when a task calls for it, the agent loads it into context. Nothing new in the model — just packaged, reusable know-how.
The key phrase is on demand. You don't pour every skill into the window at once — that bloats the context and rots it. Instead, the agent sees a short list of what's available and loads a skill's full contents only when relevant. Progressive disclosure: a lean window, capability a reach away.
This solves a real tension. You want an agent that can do many things, but can't fit all those instructions in the window at once. Skills let capability scale without the context paying for it up front. Add a hundred; the agent still loads only the one or two a task needs.
And notice what a skill actually is: mostly text — instructions the model reads — plus maybe a tool it already knows how to call. It's the prompt-is-the-program idea, packaged and reusable. Write it once, and any agent with the harness can load it. Capability becomes shareable.
The trap comes in two forms. One: thinking a skill upgrades the model — it doesn't; it's packaged context and tools. Two: loading them all to be safe — which puts you back to a bloated, rotting window, the problem skills exist to solve. The whole value is loading only what's relevant.
So a skill is a reusable capability the agent loads on demand — instructions and tools, packaged, pulled in only when relevant. That extends what the harness can do. But skills add capability; they don't control the loop. How do you inject your own logic into the cycle? Next: hooks.