I once read a great piece about making an agent lay out its plan before it writes code. Nodded. Felt obviously right. Closed the tab.
Three days later, two in the afternoon, in a rush, I typed "just do this" and let it run straight through. Forty minutes later I opened the result: it had done five times more than needed and touched three places I never expected. I knew I should make it plan first. I just didn't do it. The gap between those two — knowing and doing — is exactly what this handbook exists to close.
01What you have to remember yourself will eventually slip
Here's the blunt truth about operating knowledge: a principle whose application depends on your willpower in the moment is a weak principle. Not because it's wrong. Because you — a human — aren't always sharp, free-handed, and disciplined enough to recall it at the precise moment it matters.
You remember on the day you read the piece. You forget on the day the deadline is breathing down your neck, your boss just pinged, and the agent is sitting there waiting for an instruction. And the day you forget is exactly the day that principle was worth the most.
That's why people who've done this a while don't try to remember harder. They do something smarter: they take the principle out of their own head and wire it into the system — so the right thing no longer depends on the memory of one tired person.
02A skill is a principle with the remembering built in
A skill, here, isn't a clever tip. It's a principle that's been frozen into something installable. You install it once, and from then on the right behavior just happens — no re-instructing, no remembering.
Take my trap above. The principle is "make it plan first." Installed as a skill, it becomes a line sitting in the agent's rulebook: before editing code, always lay out the plan and stop for my go-ahead. Now I don't have to remember. The file remembers for me. However rushed I am, however curt my "just do this," the agent still stops and presents a plan — because the right thing lives in the rules now, not in my memory.
Step 2 — Freeze: you distill the principle into a concrete artifact — a rule line, a prompt, a saved command. Written once, never rewritten.
Step 3 — Install: that artifact goes into the agent's setup. From now on the right thing runs every time, even on the day you forget it entirely — because it no longer waits on your memory.
The principle layer teaches you to <i>see</i> the problem. The skill layer makes that seeing <i>happen on its own</i>. A good handbook gives you both — and, crucially, the bridge between them.
03Four shapes, one idea
Every skill in this handbook does the same thing — turns a principle into a reflex — but wears one of four jackets, depending on where you want that reflex to live.
A line of law pasted into the agent's config file. Applies to every session, forever, no reminding. For behavior you want always-on.
A block you drop into the chat right when you need it. No setup, works with any agent. For what you need occasionally, in the moment.
A multi-step routine saved behind a short name. For what you do over and over in the same sequence.
A bundle of capability the agent loads when it meets the right kind of task. For a whole skill you want the agent to reach for itself at the right moment.
Don't get lost in which shape is "fancier." The right question is simpler: do I need this reflex <i>always</i> (rule), <i>occasionally</i> (prompt), <i>repeatedly</i> (slash), or <i>automatically at the right moment</i> (SKILL.md)?
04How to read this handbook
This handbook isn't meant to be read front to back. It's a toolbox — you open it when you hit the matching job.
The use is three beats. One: you hit a recurring pain — the agent keeps saying "done" without running it, or keeps diving into fixes before understanding. Two: find the skill that matches the pain, read it for a minute to understand what it blocks, then copy. Each skill ships an artifact to paste straight in — or a prompt that tells the agent to install it for you. Three: install it, then forget it. That's the best part: once it's in, you're freed from having to remember.
A few worth installing first, because they block the most expensive falls: make it plan first before it codes, verify by running it instead of trusting "done", and diagnose from the symptom before fixing blind. Browse the whole handbook to pick the ones that match your work.
05Understand to nod, install to not forget
The rest of the garden teaches you the why — why the agent falls here, why that gate saves you. That why matters, because without it you install skills like charms, and a charm in the wrong place backfires.
But once you understand, don't stop at nodding. The nod will evaporate on exactly the day you need it. Take the principle you just grasped, freeze it into one installable line, and wire it into the system — so next time, when you're most rushed and most forgetful, the right thing still just happens. A principle in your head is a promise. A principle installed into your setup is a habit you can't miss.