05 / Skills

Grow your own skills

The shipped skill library is the smaller half of what runs this household. Muse arrives with 54 stock skills. Sitting next to them are 63 custom skills written here, and the custom half is where the daily work actually lives. That ratio is the lesson: the stock library teaches the agent generic tasks, but the skills that change what your agent is worth are the ones grown out of your own repeated work.

A skill, in plain terms, is a workflow written down as a playbook the agent follows instead of improvising. Stock skills cover the generic shape of a life: booking flights, triaging email, planning trips, managing documents, tracking fitness. They are competent starting points. They know nothing about your systems, your standards, or the ways things have already gone wrong once. A custom skill is where that knowledge goes. It is a workflow that survived contact with reality, written down so the next run starts from the lesson instead of from zero.

How a task becomes a skill here

Skills are born two ways in this house.

The first is Muse's own learned-skills loop. A daily skill-review pass watches for workflows that recur and promotes them into the custom library, marked as learned. Roughly 27 of the 63 custom skills arrived that way. You do not have to write those by hand. You have to notice them, read them, and correct them, because the loop drafts from what you actually did, including your shortcuts.

The second way is deliberate authoring, and it follows a simple rule: a workflow gets written down when it recurs, when it has failed the same way twice, or when it has to be done exactly the same way every time. Four examples show the range:

Ship. One word that runs a whole closing ritual: write the narrative record first, then commit, merge, deploy, verify the result live, and clean up. It exists because "done" kept meaning different things at different steps, and work was being declared finished at the commit instead of at the verified result
Where-were-we. A re-grounding ritual for a stale chat: reconstruct what the thread was for, check which items are still open against current state, name who owns each one, and give the single next step. It exists because long-running chats drift, and returning to one cold used to cost a full re-read
Upgrades-scan. A read-only sweep of a named surface that returns a ranked list of opportunities, which the operator then disposes item by item. It turns "anything else we should fix?" from a vibe into a checklist with verdicts
Narrative-docs-update. After any coherent arc of work, the story of what happened goes into the work's own repository, in its docs folder, where later processes will actually find it. The habit exists because records written anywhere else turned out to be invisible to the systems that needed them

Notice what these have in common. None of them teaches the agent a new fact. Each one encodes a standard: what counts as finished, what gets checked before a status is claimed, where the record lives. That is the part worth stealing. Facts go stale; standards compound.

The library is maintained, not accumulated

A grown library rots if nobody tends it, and this one shows the maintenance honestly. Some skills in the catalog arrived as recoveries after a layout change and carry a standing warning to review their instructions before reuse, because the loop preserves text it does not always curate. Superseded skills are kept as redirects rather than deleted, so an old reference lands somewhere sensible instead of breaking. The practical habits that follow:

Read a learned skill before you trust it. It was drafted from your behavior, and your behavior includes the day you were in a hurry.
When a workflow changes, change the skill the same day. A playbook that describes last month's process is worse than none, because the agent will follow it confidently.
Retire skills on purpose. A redirect that says "superseded, use this instead" is a healthy library's scar tissue, not clutter.

Where to start

Start with the one workflow you find yourself re-explaining to your agent every week. Write it down as a skill, including the step where it went wrong last time. Then let the learned-skills loop draft the next few, and edit those drafts the way you would edit a new employee's first procedure note: kindly, and in ink.

Two public starting points from this household: the claude-skills-updater and claude-code-starter repositories, both public and listed in full on the resources page of this site.