spot-skill-opportunities

name	spot-skill-opportunities
description	Find opportunities for new skills.
user-invocable	true
allowed-tools	["Bash","Read"]

spot-skill-opportunities — notice when work is skill-shaped

The recognition half of "always look for opportunities to create new reusable skills" (memories/preferences.md). skill-builder is the construction half — it scaffolds a skill once one is needed. This skill is what decides that one is needed, and it fires continuously, in the middle of ordinary work, not only at a session or PR-end checkpoint.

The same recognition applies one level down: if the emerging pattern is really a dedicated, read-only fan-out worker a heavy skill's per-item step needs (not a new user-invocable workflow), hand off to agent-builder instead of skill-builder.

When this fires

• You repeat a multi-step procedure you (or another session) already did earlier — in this conversation or a prior one recorded in memory.
• You improvise a workaround, diagnostic recipe, or decision framework that is clearly general enough to recur (not a one-off fix specific to this bug).
• The user says "again", "like last time", "we did this before", "every time you...", or gives a verbal standing rule ("always do X") that isn't yet backed by a skill.
• A review round (human or bot) surfaces the same correction more than once.
• As a standing checklist item inside record-learnings (continuous), ums (explicit checkpoint), wrap-up (session end), and post-merge (PR end) — those skills defer the recognition judgment call to this one.

Procedure

1. Name the candidate pattern. State in one sentence what the repeatable unit is (the trigger, the steps, the output) — if you can't state it crisply, it probably isn't skill-shaped yet.
2. Rule out one-off. Would this plausibly recur across sessions, repos, or users of this config — or is it specific to this single task? Only the former is worth a skill. When in doubt, lean toward surfacing it anyway and let the user decide (see step 4).
3. Check it isn't already covered. Search for an existing skill, an in-flight branch, or an open PR before proposing a new one:

   grep -ril "<keywords>" skills/*/SKILL.md
   

   (see check-open-prs-before-duplicating for the PR check). If something already owns this concern, suggest extending it — or, if a PR is already building it, redirect to that PR — instead of proposing a new skill.
4. Surface the suggestion — don't build silently. Per the standing preference, propose creating the skill (or agent); don't scaffold one unasked mid-task (that would derail the task at hand). The exception: when you're already inside an explicit authoring context (mid-ums, mid-skill-builder, mid-agent-builder), hand off directly without waiting for user confirmation.
5. Hand off to the right builder. A new user-invocable workflow goes to skill-builder; a dedicated read-only fan-out worker persona for an existing heavy skill goes to agent-builder. Once the user agrees (or the context already sanctioned it), let that skill do its own extend-first check and scaffolding — don't duplicate its step 0 here.

Relationship to other skills

• skill-builder — the construction step this hands off to. This skill never scaffolds a SKILL.md itself; it only recognizes the need and routes there.
• agent-builder — the construction step for the fan-out-worker case: when the recurring pattern is a dedicated read-only subagent for a heavy skill's per-item step, rather than a new user-invocable workflow.
• record-learnings — passive and continuous like this skill, but covers the broader category of learnings (facts, tool quirks, preferences). Its "if it's a skill, create it" step is this skill's specific case.
• ums — the explicit checkpoint version. Its checklist item "did a workflow emerge that could be a new skill?" is this skill's standing version, run continuously instead of only at the checkpoint.
• wrap-up / post-merge — session- and PR-level bookends that already embed a UMS pass looking for exactly this signal; this skill is what fires during the work, so the pattern is caught before the bookend rather than only at it.
• consolidate-skills / find-overlap — the cleanup counterparts for skills that already shipped and turned out to overlap; this skill tries to prevent that overlap from being created in the first place by checking for an existing owner before proposing a new skill.
• config-ai — the explicit, user-invoked counterpart: this skill proactively notices a skill/agent opportunity mid-task; config-ai is what fires when the user directly states a capability and doesn't yet know (or care) whether it's a skill, agent, memory, hook, or gha capability.

Anti-patterns

• ❌ Proposing a skill for a genuinely one-off task with no plausible recurrence.
• ❌ Scaffolding a skill mid-task without surfacing the suggestion first — that derails the task at hand and skips the user's call.
• ❌ Skipping the existing-skill check and proposing a near-duplicate.
• ❌ Only noticing the pattern at session end (wrap-up) instead of live, in the moment when the repetition first appears.
• ❌ Staying silent when a clear multi-step recipe repeats — the standing preference is to always surface it, not to wait to be asked.