跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

build-loop

Drive a build to actually-works, near-finish-line craft by looping build → see → exercise → check → critique → rebuild over the agent's existing tools (bash build/test, headless screenshot + vision, Playwright, axe/Lighthouse) until acceptance criteria pass or an explicit stop-condition fires — no infinite thrash. ALWAYS invoke when the user says any of "tighten this build", "iterate until it passes", "self-verify the UI", "make it actually work, not just compile", "loop until the core flows pass", or "drive this to near-finish-line". Also invoke proactively after a scaffold or feature lands and before handing a draft to a human. Composition-first over existing tools, not a new program; it states its own honest bounds — objective craft only, taste self-graded, ~80% ceiling, market validation out of scope. Do NOT use to decide WHAT to build (that's ideate) or whether a design/strategy is sound (deep-dive).

nelsonwerd · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add nelsonwerd/idea-to-ship-skills --skill build-loop

技能说明


name: build-loop description: >- Drive a build to actually-works, near-finish-line craft by looping build → see → exercise → check → critique → rebuild over the agent's existing tools (bash build/test, headless screenshot + vision, Playwright, axe/Lighthouse) until acceptance criteria pass or an explicit stop-condition fires — no infinite thrash. ALWAYS invoke when the user says any of "tighten this build", "iterate until it passes", "self-verify the UI", "make it actually work, not just compile", "loop until the core flows pass", or "drive this to near-finish-line". Also invoke proactively after a scaffold or feature lands and before handing a draft to a human. Composition-first over existing tools, not a new program; it states its own honest bounds — objective craft only, taste self-graded, ~80% ceiling, market validation out of scope. Do NOT use to decide WHAT to build (that's ideate) or whether a design/strategy is sound (deep-dive).

Build-Loop — see-and-exercise iteration until a build actually works

build-loop takes a build that compiles-but-isn't-done and drives it toward near-finish-line craft by repeating one disciplined cycle — build → see → exercise → check → critique → rebuild — until its acceptance criteria pass or a stop-condition fires. It doesn't just re-read its own plan; it runs, sees, and exercises what it built, then feeds those observations back into the next rebuild.

It is composition-first: it orchestrates tools the agent already has — bash (build/test), a headless screenshot + vision to see, Playwright/headless to exercise the flows, axe/Lighthouse to check — via prose. It is not a new program, and not a reinvention of the see-and-exercise agents vendors already ship; it's the discipline that wraps them: a target, a loop, a stop-condition, and an honest ledger of what was actually checked.

This skill is the downstream partner to prompt-pack (which sequences what to build) and ideate (which decides whether to build it). It carries prompt-pack's execute-discipline (never fake a signal you didn't run) and ideate's conditional-design (a design bar only when feel is load-bearing).

When to use this

Strong triggers — invoke without asking:

  • "Tighten this build" / "drive it to near-finish-line" / "make it actually work, not just compile"
  • "Iterate until it passes" / "loop until the core flows pass"
  • "Self-verify the UI" / "see whether what I built actually renders and works"

Softer triggers — invoke proactively:

  • A scaffold or feature just landed and you want it driven from "compiles" to "works + holds a craft bar."
  • You're about to hand a draft to a human and want the machine-checkable defects gone first.

Do NOT use this for:

  • Deciding what to build or whether it's worth building → that's ideate.
  • Judging whether a design/strategy/codebase is sound or safe → that's deep-dive.
  • A one-line fix you can make and eyeball in ten seconds → just do it.
  • Validating that people want it → out of scope (see bounds below); that's the human/market handoff.

Routing tie-breaker: build-loop answers "does this build actually work and meet its craft bar?" — not "is it the right thing to build?" (ideate) or "is it sound?" (deep-dive).

What it can and cannot do (read this before you trust it)

The honest bounds are load-bearing — they're stated here, up front, so no one reads a green loop as more than it is.

  • Catches (objective, trustworthy ground truth): broken builds, failing tests, dead/half-wired flows, console errors, a control that doesn't actually change its output, missing-asset/404s, contrast and other axe-checkable a11y violations, blown performance budgets. These are machine-checkable — this is the part that moves a rudimentary scaffold toward near-finish-line craft.
  • Soft on (self-graded taste): a model screenshotting its own UI and calling it "tasteful" is the monoculture, at the visual layer. The loop drives design hard toward a stated bar and reliably flags ugly/broken/missing; it stays unreliable on genuinely good. The different-model critic (default when design is load-bearing) blunts this, but two correlated models still aren't a user — a human spot-check is the final taste gate ("looks good to two models" ≠ "a designer signed off").
  • Cannot see (the last-mile tail — plan for a human to finish it):
    • Content wrong against a source of truth it doesn't hold — e.g. a value marked "invalid" that is actually valid in the target dictionary. The build is green, the flow works, the screenshot looks fine, and the content is still wrong. Without the external oracle, the loop has no way to know.
    • Subtle correctness/security defects with no failing check — a race, an auth edge, an off-by-one that no test exercises.
    • Whether anyone wants it — "works + looks good" ≠ "people want it."
  • Ceiling: ~80% of the craft, not 100%. "Near-finish-line" is the aim, not a guarantee; the last-mile correctness/security/taste tail is the human's. Market validation is explicitly out of scope.

If a defect lands in "cannot see," the loop's job is to name it and hand it off — never to paper over it with a passing-looking screenshot.

Prerequisite — lock the acceptance criteria first

You cannot loop without a target; a loop with no criteria either thrashes forever or stops arbitrarily. Before iterating, lock:

  1. Acceptance criteria — binary, checkable statements of "done" (a machine or a scripted flow can settle each). Detail + examples in references/acceptance-criteria.md.
  2. The design bar — conditionally. If experience/feel is load-bearing to this product (cf. ideate's conditional-design forcing-function), carry a substantive design direction from the brief — target aesthetic/vibe, 1–2 reference points, key design principles, the intended "feel" — plus concrete desi
阅读完整来源 ↗