Think about how a capable human assistant actually gets better at your job. They don't hand you a catalogue of abilities and ask you to install the right ones. They watch the work, notice the patterns, and quietly develop routines: the Monday report assembles itself, the invoice chase happens on schedule, the briefing is ready before you sit down. The competence is real, but it is background — you experience it as things simply being handled.
That is the model Kavvo's skills are built on.
How a skill is born
A skill starts as friction. You ask Kavvo to do something it has no ready routine for — say, gathering numbers from three web pages every week and turning them into a one-page summary. Kavvo can do the task once, the slow way. Then, in the background, it does something more valuable: it designs a skill for it.
- It writes the steps down as a reusable capability, with the sources, formats, and checks the task needs.
- It tests the skill against the data it already handled, so the routine is proven before it is trusted.
- It files the skill away — not in a menu you manage, but in its own working memory of how your jobs get done.
Next Monday, the summary is simply waiting. The activity stream shows the run, start to finish, so the invisibility never becomes opacity.
Why not a plug-in store?
Stores have three structural problems. First, discovery: you must know a capability exists and go find it before you can benefit. Second, fit: a store skill is written for everyone's average task, not for your suppliers, your formats, your quirks. Third, upkeep: installs, updates, permissions, and breakages become your chore list.
Self-created skills invert all three. Discovery is automatic — the skill exists because your work needed it. Fit is exact — it was written around your actual task. And upkeep is the agent's problem: when a source changes or a step fails, repairing the routine is background work, reported in the stream like everything else.
Skills compound
The quiet magic is accumulation. A skill that pulls figures can feed a skill that summarises; both can feed the weekly review automation. Over weeks, Kavvo's background toolkit becomes a map of your working life — one no store could sell you, because it isn't generic. It is, quite literally, learned from the job.
And none of it asks anything of you. There is no skills page to organise, no toggles to curate. If a skill ever does the wrong thing, the activity stream shows it, the conversation can correct it, and the correction itself becomes part of what Kavvo remembers.
The principle
We think the right division of labour is this: you bring the intent; the agent builds the machinery. Skills are where Kavvo keeps that machinery — created by the agent, owned by your install, invisible until the moment they save you an hour. Which, if we have done our jobs, will be most mornings.