本文へ移動
cccskills
無料GitHub で公開

browser-app-delivery

Deliver a runnable local browser app so the user can operate it on the first handoff. This delivery skill does not choose whether a requested local system should be a browser app. On a current turn that explicitly says to implement a brand-new browser project now, load bundled:greenfield-project-scaffolding before project work only after no-peek or permitted top-level placement facts prove that no named or suitable current-folder target resolves the root; never hand off from planning or answer-only turns, existing work, a named or suitable target, or standalone or single-file delivery. Resolve user-blocking product or interface choices with an available interaction surface; resolving those choices alone never authorizes implementation. Always call read_skill for bundled:browser-app-delivery before creating or materially changing a local browser app that owns its start path or a playable browser game. Also load it when the user requests a paste-ready, single-file, offline, or immediately playable browser artifact, even if no local start path should remain, unless the user says they will open or check it themselves or explicitly declines verification. On completed local-server delivery, include exact start commands, a concrete local URL, and explicit browser-open wording. On completed standalone delivery, hand off the exact artifact path or paste URL and smoke result without inventing a server. Do not load it for a self-contained static file that has no owned start path if the user either says they will open or check it themselves or explicitly declines verification. Do not load it for component/style-only edits, deployed sites, backend/API-only work, browser QA, review, explanation, plan-only, or explicit stop/no-tool turns.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md10.0 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Browser App Delivery

For a brand-new browser project, apply this hard boundary before project work. Do NOT call read_skill for bundled:greenfield-project-scaffolding on a requirements-guidance, planning, option-selection, multi-decision, or answer-only turn; from prior-turn answers, authorization, or reminders without new implementation-now words; for existing work, server start/restart, verification, or handoff-only work; for a named path or a confirmed-suitable here/current-folder target; or for a requested standalone, paste-ready, single-file, or snippet artifact.

Only when this same user turn explicitly says to start, implement, build, scaffold, or go ahead with a new project now may unresolved placement trigger the handoff. Do not load the greenfield skill merely to decide whether it is eligible. If inspection is forbidden, do not list or read the workspace; use only the request and pwd. Otherwise inspect only top-level placement facts first. Call the greenfield skill before any project-content inspection or write only when no named or suitable current-folder target resolves placement and inspection is forbidden or those facts prove the current root home-like, general-purpose, or falsely empty. Wait for its body, resolve the root, then continue this workflow.

Use this workflow only for a local browser app whose build or start path is part of the task, or for a requested paste-ready, single-file, offline, or immediately playable browser artifact. Current user instructions always win.

Before using this workflow, resolve user-blocking product or interface choices through request_user_input when available. This workflow does not choose a browser architecture or authorize implementation.

Product correctness precedes delivery closure. Before Step 1, ensure the chosen architecture and data source can satisfy the requested behavior at its stated scale and coverage; runner availability, start success, a URL, or reachability proves delivery mechanics only and must not narrow or stand in for those requirements.

  1. Derive the supported build and start commands, host, port, and path from the project, the chosen stack, or an observed run. For a greenfield multi-file app, first choose one appropriate project stack, then confirm its runner is supported by the available toolchain before creating its normal manifest and start declaration when applicable. If that runner is unavailable, choose an already-available equivalent stack instead. Successfully run that exact start command. Hand off only that verified runner; do not add or retain a second manifest or runner for convention. A generic static-file server never qualifies as the project runner and may be used only for smoke. An explicitly standalone artifact skips this requirement. Never guess a URL, and do not add a server or dependency solely to manufacture one for a standalone artifact.
  2. Preserve live user processes and persistent data. Prefer the existing dev server and hot reload. Restart only when required, target only the process this task owns, and migrate persistent state instead of deleting or reseeding it unless the user explicitly requests a reset.
  3. Verify changed behavior once with the smallest suitable check. A successful HTTP response is enough to establish reachability. Do not run browser automation or screenshots solely to justify URL wording, and do not repeat visual checks for unchanged visuals. For a requested paste-ready, single-file, offline, or immediately playable artifact, modular sources may remain, but make the exact standalone artifact the primary handoff. Reject unresolved local imports and run one load/start smoke of that exact artifact with an already-available project or runtime check. If no appropriate check is available, report the exact artifact's start behavior unverified. HTTP success proves only reachability. Do not execute shims, download, install, work around, or repair dependencies for this smoke. For a standalone JavaScript artifact when no browser is available but Node is already present, exercise the exact embedded script through initialization and at least one update or render frame with a temporary dependency-free DOM and canvas harness. Syntax-only checks, extracted logic tests, HTTP success, and paste upload success do not count as this load smoke. Fix every runtime error and repeat the smoke once. This temporary test harness is not a package or runtime shim. Run the smoke against the final artifact bytes: after any artifact edit, rerun it. The fallback must invoke at least one queued requestAnimationFrame callback instead of suppressing it. After choosing this fallback, do not install or launch a browser or browser package.
  4. Close browser delivery before handoff. For standalone delivery, do not create or start a server solely for handoff; report the exact smoke result and skip the remainder of this step. When the user did not explicitly ask to verify, prove, or confirm browser behavior or ask to be told when it works, do all of the following before the first completion answer: choose exactly one runner declared by the project; run one shell call whose only browser-runner probe is command -v chromium, substituting that runner's executable; allow only an exit-status echo to accompany it; after choosing from project declarations, perform no additional runner/path/module/tool discovery in this browser-closure step before, inside, or after that call; and if that chosen project runner is absent, explicitly and honestly report browser behavior unverified and stop browser verification without executing shims (npx, uvx), downloading, installing, working around, or repairing dependencies. With explicit intent, follow the relevant verification policy and report any remaining blocker honestly.
  5. Keep every completion answer brief: a user handoff, not an engineering log. This changes only the report, never implementation or verification. Lead with the outcome and strongest verification. Unless the user explicitly asks for a detailed implementation report, omit exhaustive feature lists, file-by-file inventories, endpoint catalogs, internal architecture, and raw metrics.
  6. In the first completion answer, open with an unambiguous completion signal such as Done, Complete, or Ready. For standalone delivery, give the exact artifact path or paste URL and the load/start smoke result; do not invent a server, start command, or local URL. Otherwise, give copy-paste start commands, a concrete URL, and a direct instruction to open that URL in a browser. A concrete URL must be one the user can open from their own machine. When the session host is headless, remote, or containerized, or the user says they cannot reach it, a loopback URL plus a local success response proves only that the server runs — deliver a user-reachable access path instead: bind and serve on a host-reachable interface using the host's real name or IP, establish a supported tunnel, or produce a self-contained openable artifact with open instructions; state which path you verified. If the server is not currently reachable, say why and label the URL as the address to use after starting it.
  7. State current runtime status honestly. Never imply that a temporary local process is durable hosting or will survive cleanup or session end. Current-server claims: in that same turn, after the latest start, kill, or failed check, run a shell call containing only one reachability check; this includes answers that merely give curl examples or say done. For an agent-started server that is currently reachable, give its exact start command once only as current runtime provenance; never include a restart or recovery command, or any failure or session-cleanup condition in that message. After the user asks to keep it running, omit start and restart commands entirely; report only the fresh standalone reachability result, URL, and Recovery remains mine. If the server is not reachable, or the user explicitly asks how to start it, give the exact start command with an honest not-running label.

Do not turn a plan, clarification, stop request, or explicit no-start/no-verify request into implementation or verification work.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

3d-object

無料

three.js model, downloadable as OBJ or GLB

日本語の概要は準備中です。原文の説明を表示しています。

asgeirtj/system_prompts_leaks7万2026年10月12日 更新

Manage your private to-do list for the user. Use to add or update tasks, record what they are waiting on, track blockers, mark tasks complete or canceled, and clean up the list. Not for Dreamers or external task trackers.

日本語の概要は準備中です。原文の説明を表示しています。

asgeirtj/system_prompts_leaks7万2026年10月12日 更新

agents

無料

Run a goal as a project: you coordinate. Use for "agents <task>", "take this to done", "work on this in parallel", "what are my threads doing", "pick up <slug>", "resume <slug>", "keep going until it is merged". Under `/agents` you do the work yourself unless it needs several lanes or a long wait — you judge; the plan line says why. No target you could verify (a file, API, number or measure) → the grill interview first, not threads. Read this skill before proposing anything for `/agents`: without it a proposal is in-session subagents.

日本語の概要は準備中です。原文の説明を表示しています。

asgeirtj/system_prompts_leaks7万2026年10月12日 更新

Timeline-based motion design

日本語の概要は準備中です。原文の説明を表示しています。

asgeirtj/system_prompts_leaks7万2026年10月12日 更新

The user's synced Apple Health (HealthKit) data: daily metrics (steps, distance, calories, heart rate, HRV, VO2max), sleep sessions (stages, quality, efficiency), and workouts.

日本語の概要は準備中です。原文の説明を表示しています。

asgeirtj/system_prompts_leaks7万2026年10月12日 更新

Create, read, edit, or manipulate Word documents (.docx) and Word templates (.dotx). Use whenever a build task's artifact kind is document with the default docx output, or the task mentions a Word doc, .docx, or .dotx, extracts or reorganizes content from one, inserts or replaces images, does find-and-replace in one, or works with tracked changes (redlines) or comments. Covers python-docx generation, raw OOXML editing of existing files, document structure and formatting, and render verification. Not for PDFs, spreadsheets, or Google Docs.

日本語の概要は準備中です。原文の説明を表示しています。

asgeirtj/system_prompts_leaks7万2026年10月12日 更新

asgeirtj のスキルをすべて見る

このスキルの問題を報告する