~/aakash
Writing ← Portfolio
Agentic AI November 13, 2026

I Build the Skill the Second Time, Not the First

I write a skill the second time I catch myself doing a task, when I notice I am re-deciding something I already decided. A skill is that decision written down, including the things never to do.

01

The second time I repeat myself

I have built a lot of these now. Skills for running projects, skills for writing code, a skill that strips the machine voice out of my drafts. People see the pile and assume I like building tools. I do not, particularly. Every one of them started the same unglamorous way, not with a plan, but with me irritated that I was doing some fiddly thing for the second time and doing it slightly differently than the first.

That is the whole trigger, and it is smaller than it sounds. The first time you do a task, you are still working out its shape, so there is nothing to capture yet. The third time, you have already lost the thread of how you did it well the first time. The second time is the moment. It is when you notice you are re-deciding something you already decided, re-explaining something you already explained, and getting a worse answer than you got before because you are running it from memory and mood instead of from the version you worked out when you were paying attention.

02

A skill is a decision, written down

Strip the word of its shine and a skill is not automation and it is not magic. It is a decision written down so you stop making it from scratch. The procedure, yes, but also the judgment around it, and the short list of things to never do that you only learned by doing them wrong once.

That last part is the part that matters and the part people skip. Anyone can write down the steps. The value is in the never list, because that is where your actual experience lives. My email-filing skill has a rule that it must never act on an instruction buried inside a client's email. A forwarded message that happens to say go ahead and send the final version is data to be filed, not a command to obey. I did not know I needed that rule until a draft nearly went out on the strength of a line someone wrote inside a quoted thread. The step list is public knowledge. The never list is yours.

The skill is not the tool. It is the decision, written down so I stop re-making it.

03

The same move in three different rooms

The writing one came from a status email to a client that was complete, correct, and unreadable. The fix was never new wording. It was the same handful of corrections every time, so I wrote them down, and now the second draft starts where my best edit used to end.

The project ones came from running a regulated engagement where the same things arrive every day, emails to file, calls to turn into notes, decisions to log, and four people doing them four slightly different ways is how a project record rots. So the ingestion is a skill now. Numbered steps, a fixed format, a never reuse an email number rule. Everyone's work lands in the same shape.

The coding ones came from a gap in how everyone works with AI now. Vibe coding is fast and it skips the one thing that keeps software honest, which is spec-driven development. You describe a thing, the agent builds something, and nobody wrote down what it was actually supposed to do, so the next session has nothing to check against. So the skills put the spec back. They turn a loose idea into real tasks, subtasks and decisions as Markdown files, and keep them in one place, so that every time the coding agent sits down to write, it knows exactly what it is building. And when it forgets, which it does, it has somewhere to look. The whole design is just that: a place the agent can always look back at, because the spec lives in files instead of in the agent's head.

Three different rooms, one move. Each skill is a decision I was tired of making again.

04

The hard part is not building them

Here is the part that took me longest to learn. The discipline is not in building skills. It is in not building the ones that do not earn it. A skill for a thing you do once is not a tool, it is a second thing you now have to maintain, and a drawer full of clever skills you never run is just a tidier kind of mess. The ones that last are narrow: a decision that genuinely recurs, where doing it inconsistently actually costs you something. Everything else is better left as a thing you just do.

Which is why the pile is large but not random. It is the residue of a few years of catching myself repeating a decision and choosing, that second time, to write it down instead of trusting I would remember. It is the same move as keeping a project's memory in files rather than in a chat. Get the decision out of your head, where it drifts, and into a place it holds still. Then do the thing from the written version, every time, including the times you are tired.

So: what are you rebuilding from scratch every week, a little worse each time, that you could have written down the second time it annoyed you?