Skillify

Capture the work you just did as a reusable, parameterized skill — then register it in the resolver so it can be found and reused. This is the move that compounds a one-off into a capability.

"You just do anything, and then I like it, I say skillify it — and it becomes basically like a tool call or a method call. And the most important part is plugging it into the resolver." — Pete Koomen / Andrej Karpathy, YC

When to run

• The user says "skillify this" / "make this a skill" / "save this as a skill".
• You just completed a multi-step procedure that you (or a teammate) will plausibly do again.

Do not skillify trivial one-liners or anything a single existing tool already does.

Procedure

1. Identify the procedure. Look at what was just done in this session — the steps, the order, the decisions, the gotchas. Strip out the one-off specifics (this particular file, this particular value).

2. Generalize into parameters. Anything that changed between "this run" and "the next run" becomes an input, not a hardcoded value. A good skill is the shape of the work, not a replay of one instance.

3. Check before you write — DRY. Read the resolver (see "The resolver" below) and skim existing skills. If one already covers this, extend it with a parameter instead of adding a near-duplicate. Adding skill #2 that overlaps skill #1 makes the library worse, not better.

4. Write the skill. Create skills/<kebab-name>/SKILL.md with frontmatter and a tight procedure:

   ---
   name: <kebab-name>
   description: <what it does> + Use when <the trigger>.
   ---
   
   # <Title>
   
   ## Inputs
   - `<param>` — <what it is>
   
   ## Procedure
   1. ...
   2. ...
   

   Keep it minimal. The procedure should read like the steps a careful teammate would follow — no speculative options, no abstractions for a single use.

5. Register it in the resolver. Add one line to the index the agent actually reads: name + a one-line use-when + a link to skills/<kebab-name>/SKILL.md. A skill that isn't in the resolver doesn't exist.

6. Verify resolvability. Run check-resolvable to confirm the library is still DRY and MECE after the addition — no duplicate, no overlap, no gap.

The resolver

The "resolver" is whatever flat index your harness reads to discover capabilities — AGENTS.md / CLAUDE.md skills list, a tool registry, the skills/ directory itself. Treat it as the source of truth: every skill is reachable from it by name, with a one-line description that says when to use it. Register on create; prune on merge.

Output

State plainly: the new skill's path, the one-line resolver entry you added, and whether check-resolvable flagged anything to merge.