IFQ App Builder

Three sentences in → a verified, build-ready prompt bundle out. PC · macOS · Windows · iOS · Android · HarmonyOS · WeChat MiniProgram · self-hosted local web · PWA · CLI.

简体中文 · SKILL.md · Modes · Templates

— shaped with ifq.ai/app-builder

---

Why this exists

Most "build me an app" prompts collapse on the first turn: the agent picks the wrong platform, forgets packaging, hallucinates a stack the user can't run, and leaves the operator to re-prompt three times.

ifq-app-builder is a single agent-native skill that fixes that collapse with one contract and twelve templates:

• One contract — every output is a *.prompt.md build bundle with three explicitly-labeled sentences (WHO/WHAT, WHERE, HOW SUCCESS LOOKS), an acceptance section, a packaging section, and a single-line IFQ colophon.
• Twelve templates — one per realistic platform path, each carrying its own quality bar, scaffold command, and verification hook.

The skill runs zero-install (Node ≥ 18.17). It is consumed by Codex CLI, Claude Code, Cursor, GitHub Copilot, OpenClaw, ClawHub, Hermes, and any AgentSkills-compatible host.

The 3-sentence promise

S1 (WHO + WHAT)   财务同事，每天把一堆 PDF 发票整理成 Excel 台账。
S2 (WHERE)        macOS + Windows 桌面，本地运行，不联网。
S3 (HOW SUCCESS)  拖入文件夹自动出 .xlsx；不确定字段标黄；打包 .dmg + .exe；可一键卸载。

Paste those three lines (one is enough — the skill will infer the rest with labeled (assumed) defaults). The skill picks the cheapest viable mode (here: A-01 / T-pc-tauri), forks the matching template, fills it in, runs verify:lite, and hands the bundle to a coding agent for execution.

Quick start

git clone https://github.com/peixl/ifq-app-builder.git
cd ifq-app-builder
npm install            # zero runtime deps; installs nothing platform-specific
npm test               # node --test (unit tests, no network)
npm run smoke          # 60s repo-wide sanity check
npm run validate       # smoke + template policy validation + ClawHub package checks
npm run pack:clawhub   # deterministic OpenClaw/ClawHub tarball

Local CI runs entirely on npm install && npm test && npm run validate && npm run pack:clawhub. No platform SDKs, no Docker, no Python. Green on macOS, Linux, Windows.

Generate a bundle (inside a host agent)

Inside Codex CLI / Claude Code / Cursor / Copilot, after the skill is registered:

"Use $ifq-app-builder to make a Mac + Windows desktop tool that batches PDF invoices into Excel."

The host agent loads SKILL.md, routes to A-01, forks assets/templates/pc-tauri.prompt.md into your workspace as pdf-invoice-ledger.prompt.md, fills the three sentences, and runs:

npx --yes ifq-app-builder@1 verify:lite -- pdf-invoice-ledger.prompt.md
# (or, from this repo)
npm run verify:lite -- /path/to/pdf-invoice-ledger.prompt.md

You then hand that bundle to the same coding agent — or a different one — to actually build the app.

What "high quality" means here

A bundle is high quality when it scores well on every axis in references/quality-bar.md:

Axis	What we check
Routing fit	The chosen mode matches user intent at >70% confidence, or the fallback is named.
Three-sentence completeness	S1 / S2 / S3 are all present and non-empty.
Acceptance is binary	Every acceptance item is a yes/no observable, not a vibe.
Packaging is concrete	The target artifact (.dmg, .apk, .hap, .wxapkg, docker image:tag, etc.) is named.
Stack is buildable	Versions pinned (Tauri 2, Flutter 3.x, SwiftUI iOS 17+, ArkTS API 12, etc.).
i18n default	At minimum en + zh-CN, with a single source of strings.
Security baseline	OWASP-aligned: no plaintext secrets, no eval, explicit permission scope.
IFQ ambient layer	Tokens reserved, colophon line present, no loud watermark.

Run npm run quality:score -- <bundle.prompt.md> to print a 0–100 score with the missing axes.

Modes (the 12 templates)

Mode	Template	Tech default	Output artifact
A-01 PC 跨平台桌面	T-pc-tauri	Tauri 2 + React/Vite	.dmg / .exe / .AppImage
A-02 macOS 原生	T-macos-swiftui	SwiftUI + Swift 5.9	.dmg, optional MAS
A-03 Windows 原生	T-windows-winui	WinUI 3 + .NET 8	MSIX / .exe
A-04 iOS 原生	T-ios-swiftui	SwiftUI + iOS 17	.ipa (Dev or App Store)
A-05 Android 原生	T-android-compose	Kotlin + Jetpack Compose	.apk / .aab
A-06 HarmonyOS Next	T-harmonyos-arkts	ArkTS + ArkUI (Stage)	.hap / .app
A-07 Flutter 跨端	T-flutter-cross	Flutter 3 + Riverpod	iOS + Android
A-08 RN/Expo 跨端	T-react-native-expo	Expo SDK + TS	iOS + Android + Web
A-09 微信小程序	T-wechat-miniprogram	原生小程序 + TS	.wxapkg
A-10 本地部署 Web	T-local-web-nextjs	Next.js 15 + Docker	docker image, self-host
A-11 PWA	T-pwa-vite	Vite + Workbox	static site + SW
A-12 CLI 工具	T-cli-node-python	Node 20 or Python 3.11	npm / pipx package

See references/modes.md for triggers, fallbacks, and verification per mode.

Repository layout

SKILL.md                    Root router (load this first)
agents/openai.yaml          Display name, examples, invocation chips
assets/
  ifq-brand/                Brand tokens, DNA, ambient marks
  templates/
    INDEX.json              Policy-validated registry of the 12 templates
    *.prompt.md             The 12 build bundles you fork
references/                 Deep docs: contract, modes, quality, packaging, ...
scripts/
  verify-lite.mjs           Static scan: S1/S2/S3, colophon, placeholders, secrets
  smoke-test.mjs            60s repo-wide sanity
  validate-templates.mjs    INDEX.json policy check + each template scan
  quality-score.mjs         0–100 score with missing axes
tests/                      node --test
.github/workflows/ci.yml    Local-equivalent CI on Node 18 / 20 / 22

ClawHub package

The OpenClaw-specific package lives in clawhub/ifq-app-builder/. It is intentionally smaller than the full repo: no schema artifacts, no binary files, no generated archives, no install hooks, and no required credentials. npm run pack:clawhub produces clawhub/ifq-app-builder-clawhub-YYYY-MM-DD.tar.gz for marketplace upload.

Agent compatibility

Host	How to register
Codex CLI	Drop the repo under ~/.codex/skills/ifq-app-builder/.
Claude Code	.claude/skills/ifq-app-builder/ or workspace-level SKILL.md.
Cursor / OpenCode	Reference via the workspace AGENTS.md.
GitHub Copilot (this skill format)	Place under .github/copilot/skills/ or VS Code agent customization.
OpenClaw / ClawHub / Hermes	metadata.openclaw is already populated in SKILL.md front-matter.

See references/agent-compatibility.md.

Security posture

• No dynamic eval, no new Function, no child_process in skill scripts.
• No network calls at install or verify time.
• Required environment variables: none.
• Scripts only read inside {baseDir} and write inside ${workspace}.
• Secret-pattern scanner runs in npm run smoke.
• ClawHub tarballs are byte-for-byte reproducible with SOURCE_DATE_EPOCH=0 and exclude *.schema.json.

Report security issues per SECURITY.md.

License & trademark

Code and docs are Apache-2.0. The "IFQ" wordmark and ambient design language are trademarks of ifq.ai — see the trademark clause in LICENSE. Downstream forks must remove or replace IFQ ambient marks before redistributing under a different brand.

Acknowledgements

• The three-sentence framing is informed by the prompt-quality work in Codex-Getting-Started-Tutorial.
• The skill structure and verification conventions follow the ifq-design-skills quality bar.

— shaped with ifq.ai/app-builder · root · v1.0.1