skillify — fast-capture entrypoint

skill-creator already has everything needed to capture a workflow from conversation history (its "Capture Intent" step) and a lite path for when formal evals aren't warranted (its "just vibe with me" escape hatch). What's missing is a one-word trigger that jumps straight there without first explaining to skill-creator's full interview flow that this is a quick capture, not a from-scratch skill design session.

skillify is that trigger. It is deliberately thin: draft, confirm, write, hand off. It does not run evals, benchmark variance, or reimplement any part of skill-creator's interview loop. If you catch yourself doing that inside this skill, stop — that logic belongs in skill-creator, not here.

When this fires

• "skillify this"
• "capture this pattern"
• "turn this into a skill"
• "make this repeatable" / "we'll want to do this again"
• Any moment where the user points at something just done in this session and asks for it to become a reusable skill

What it does

1. Extract the pattern from conversation history. Look back over the turns that cover the workflow being referenced (usually the last 10-30 turns, or however far back the user points). Identify:
   • The trigger: what the user said or the situation that kicked this off
   • The steps actually taken: tool calls, decision points, any corrections the user made along the way
   • Inputs consumed and the output produced
   • Anything that varied by judgment rather than a fixed rule (a candidate edge case)
2. Draft a SKILL.md — don't write it to disk yet:
   • name: short, kebab-case, matching the pattern's action
   • description: trigger phrases plus what it does, specific and a little "pushy" (skill-creator's term — bias toward firing rather than under-triggering)
   • version: 1.0.0 for a new skill
   • allowed-tools: only the tools the extracted pattern actually used
   • Body: numbered steps of the workflow, in imperative form
   • An explicit Edge Cases section listing anything the history couldn't resolve — gaps to ask the user about, not guesses to bake in silently
   • If a step verifies/validates against a flaky or slow external resource (a third-party API/service, an OS-integration shell-out), check docs/skill-authoring.md rule 3 before drafting it — that step must run through run_bounded_external (skills/self-verify/references/bounded-external-waits.md), not an unbounded wait.
3. Show the draft and confirm before writing. This is the safety step — never skip it. If the user wants changes, iterate on the draft in conversation. Only write to disk once they say to proceed.
4. Write and register. Once confirmed, write to skills/<name>/SKILL.md in the target plugin. If the target is this pfeff/claude-skills marketplace's mbp plugin, see "Registration" below — a skill directory alone does not make it resolvable as a slash command. For any other repo, check how a recently merged sibling skill was wired in before assuming the same shape applies.
5. Hand off to skill-creator when the pattern warrants it. A pattern warrants the fuller loop when it has real edge cases, ambiguous triggering, or the user wants it tested at scale before trusting it. Say so explicitly and let the user decide whether to continue into skill-creator's interview/eval/benchmark loop now or later. Skip the handoff for a genuinely trivial one-off — the drafted SKILL.md from step 2 is often enough on its own.

Registration (this marketplace)

Adding a skill directory under skills/ is not sufficient for it to resolve as a slash command in this repo. It must also be added to the skills array of the mbp plugin entry in .claude-plugin/marketplace.json (alphabetical). Skipping this step is a real failure mode in this repo: PR #133 had to retroactively register six skill directories that had sat on disk for a while with no entry in that array, so none of them resolved as slash commands despite looking complete.

Bump the version in this same PR. Run scripts/bump-version.sh — it minor-bumps version in plugin.json and metadata.version in marketplace.json together, keeping them in sync. The bump is atomic with the change that necessitates it; see docs/skill-authoring.md §2. The old doctrine kept the bump out of the PR to dodge byte-identical concurrent collisions; those are now caught loudly by the version gate (.github/workflows/version-gate.yml), which fails any PR whose version is not strictly greater than main's current version — so if a sibling PR bumps first, rebase and re-run the script.

What this is NOT

• Not a reimplementation of skill-creator's interview, eval-writing, or benchmarking — hand off for that.
• Not a place to accumulate skill-authoring best practices — those live in skill-creator's own references.
• Not a bypass of the confirm-before-write step, even for "obviously fine" drafts.

See also

• skill-creator (vendored, not owned by this marketplace) — the full draft → test → benchmark → iterate loop
• lessons-learned — retrospective analysis of what went wrong; use that when the goal is diagnosing a problem, not capturing a working pattern