Before you build another agent skill, check the skills you already own and get one evidence-based verdict: EXTEND, COMPOSE, or BUILD NEW.
- What problem does it solve? It exposes forgotten overlap before duplicate agent skills are built with competing trigger surfaces.
- What does it return? Exactly one short verdict: extend one existing skill, compose two existing skills, or build a genuinely separate skill.
- How does it compare candidates? By the problem solved, input/output shape, and judgment-versus-execution nature — never by names or keyword similarity.
- Does it create or edit skills? It checks reuse first, then hands off to the authoring workflow and continues the creation or update already authorized. No extra approval is needed for a routine reuse choice.
- Can the check be overridden? Yes. “Build it anyway” ends the gate in the same turn without relitigating the decision.
A skill you wrote and used twice is, six months later, indistinguishable from a skill you never wrote. Libraries grow past what their owners remember, so the same problem quietly gets solved twice — or a near-duplicate gets built with a slightly different trigger surface, and now two skills raise their hand for the same request. This gate runs at the moment of "make me a new skill for X": it inventories what already exists, compares by what each skill does rather than what it is called, and hands back a one-screen verdict.
- Inventories installed skills and your skill directories at the description level — bodies are opened only for the nearest 1–3 candidates.
- Compares by core function on three axes: the problem solved, the input/output shape, and judgment-vs-execution. Name and keyword similarity is banned as evidence in both directions.
- Returns exactly one of three calls:
- EXTEND — which skill, which section, what change.
- COMPOSE — which two existing skills together, with the combined usage.
- BUILD NEW — overlap ruled out, plus a drafted description boundary clause against the nearest neighbor, so the next collision is prevented too.
- Quotes only the description fragment needed to show functional overlap, keeping the check short.
- Steps aside instantly when overridden: "build it anyway" ends the check, no relitigating.
- "Make me a skill that turns my meeting notes into action items."
- "I want a new skill for X."
- "Build me an agent that reviews my PRs for style."
- 「做一个 skill,帮我整理会议记录」
- 「帮我建个 agent 做数据清洗」
And when it doesn't: requests to use an existing skill just get routed; ordinary product or feature ideas are handled by the relevant task workflow; and "直接新建别查了" / "skip the check" is respected silently — the inventory doesn't run behind your back.
npx skills add m1nga/extend-firstAuthored fixture: suppose the reachable catalog contains doc-digest, which
extracts and condenses PDFs, and brief-format, which formats supplied summaries
as one-page briefs. These names and capabilities are illustrative.
You: Make me one skill that turns a long PDF into a one-page decision brief.
Agent: COMPOSE — reuse doc-digest for extraction and brief-format for the final page. I will continue building the requested single entry point using those two capabilities, with a clear input/output handoff and missing-dependency recovery, through the authoring workflow.
If neither candidate is installed or reachable, the verdict is BUILD NEW within the inspected scope, with the catalog limitation disclosed. It does not claim that no alternative exists anywhere. An ordinary request to summarize a PDF does not activate this creation check.
- idea-probe — the sibling gate for everything that isn't a skill: product, site, and feature ideas get persona wind-tunneling there. If your idea is a skill, extend-first checks the shelf first; if it survives as BUILD NEW and you want the concept itself tested, idea-probe is next.
- scenario-probe — after a BUILD NEW verdict, the drafted boundary clause and the eventual description should be wind-tunneled before release. extend-first prevents overlap at birth; scenario-probe catches what slipped through.
Three observations shaped this skill:
- One- and two-use skills are forgotten skills. The gate exists because human memory of a library decays faster than the library grows — which also means the gate's value increases with every skill added. At five skills it is ceremony; at twenty-five it is the only thing standing between you and a shelf of near-duplicates.
- A wrong merge costs more than a duplicate. A duplicate wastes a directory. A bad merge bloats a working skill's trigger surface until it misfires for everyone, and unpicking it costs more than either skill did. That asymmetry is written into the verdict logic: uncertain calls go to BUILD NEW.
- A gate that can't be overruled becomes a gate people route around. The veto is one phrase and takes effect the same turn. The gate's job is to make overlap visible, not to win arguments — a check the user trusts to lose gracefully is a check they'll keep running.
Before release, this skill went through an 8-scenario wind tunnel — 5 personas, 2 languages, including a fresh-install stranger with an empty shelf and head-to-head trigger collisions against 21 co-installed skills. Score: 3 clean passes, 5 degraded-risk findings, 0 harmful fires, 2 correct silences. Every finding below shipped as a fix before v1.
Messy dictation, Chinese (pass): "那个,帮我搞个 skill 吧,就下载完自动把文件夹里乱七八糟的东西归归类那种" — the gate opened, compared by core function against the nearest neighbor, and returned a one-screen BUILD NEW with a drafted boundary clause. No ceremony, one screen, veto intact.
Correct silence: "Probe this idea for a skill that auto-files my receipts" — three skills could have raised a hand here. extend-first stayed down: the verb is probe, not create, so the request belongs to idea-probe. A gate that fires on everything teaches you to stop consulting it.
Caught in the tunnel: on "make me a skill…", the stock skill-creator claims the same sentence — and a model that jumps straight to authoring skips the duplicate check entirely, which is this skill's whole reason to exist. The earlier description stated its position explicitly: runs BEFORE skill-authoring tools — verdict first, then build. The probe also caught the empty-shelf edge: a reachable-but-bare shelf is now an instant one-line BUILD NEW, never a nag for directory paths.
Probe method: scenario-probe
Built and maintained by Ming. The design notes above explain the real problem and tradeoffs that shaped this skill.
An independent agent simulation checked a scoped usage scenario after the instruction cleanup. This checks instruction behavior, not human adoption or measured time savings.