Communitygithub.com

build-project

Shows an agent how to art-direct and build "build project" so the concrete output is a landing page, homepage, or portfolio. The procedure moves through Build a project from idea to running app (universal), When to use / not, Adapt to the project (fill once). Use it when a new assignment calls for this craft.

build-project 是什麼?

build-project is a Claude Code agent skill that shows an agent how to art-direct and build "build project" so the concrete output is a landing page, homepage, or portfolio. The procedure moves through Build a project from idea to running app (universal), When to use / not, Adapt to the project (fill once). Use it when a new assignment calls for this craft.

相容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/BBguss/nouvem-agent-skills/tree/HEAD/build-project

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

Build a project from idea to running app (universal)

Origin: repeated greenfield builds across very different project types. Model-agnostic and harness-agnostic; neutral terms (sub-agent, background, heartbeat, isolation) follow task-delegation §0.

When to use / not

  • Use: a new project, a new app inside a monorepo, or a rebuild from scratch.
  • Idea still vague (goal, users or approach undecided): run idea-shaping first and build from its approved brief.
  • Do not use: a change to an existing codebase (code-change or plan-then-code); only getting an existing repo to run (local-dev-setup); diagnosing a problem (analyze-problem).
  • Read-only words in the request ("analyse", "just propose") mean: produce the plan of section 1-2 only (mode-gates).

Adapt to the project (fill once)

PlaceholderMeaningExample
<PROJECT_TYPE>one of the reference types below, or a mixsaas + landing-page
<STACK>chosen stack and whyNext.js + Postgres + Prisma
<PKG_MANAGER>package manager, pinnedpnpm 9
<DEPLOY_TARGET>where it will runstatic host, container, VPS, serverless
<TEST_CMD> / <LINT_CMD> / <TYPECHECK_CMD>verification commandspnpm test, pnpm lint, tsc --noEmit

Type references (load only the one(s) that match)

TypeFile
Portfolio / personal sitereferences/portfolio.md
Landing page / marketing sitereferences/landing-page.md
CMS / blog / content sitereferences/cms-blog.md
SaaS (auth, tenants, billing, roles, onboarding, email)references/saas.md
Admin panel / dashboardreferences/admin-panel-dashboard.md
E-commerce / shopreferences/ecommerce.md
REST / GraphQL API, backend servicereferences/api-backend.md
Internal tool / back-office / script UIreferences/internal-tool.md

Each reference holds: feature checklist, data model sketch, 2-3 stack options, security and SEO/a11y musts, definition of done.

Steps

1. Clarify (ask only what changes the outcome)

Ask at most 3-5 questions, one message, each with a recommended default so the user can answer "defaults". Ask only when the answer changes stack, data model, cost or scope:

  • Who uses it and how many (public visitors, signed-in users, staff only)?
  • Content editors who are not developers? (decides CMS vs files in repo)
  • Accounts, payments, multi-tenant, i18n, offline/mobile?
  • Hard constraints: existing hosting, language the team knows, budget, deadline.
  • Must-have features for v1 versus later. Do not ask about colors, fonts, folder names or library minutiae; decide and state them. If the user already answered, do not ask again.

2. Choose the stack

  • Start from the reference's options; prefer the simplest one that meets v1 (static before server, one service before many, managed DB before self-hosted).
  • Respect what the team or existing infrastructure already uses over fashion.
  • State the choice in one short table: layer, choice, reason, alternative. Record it as a decision (living-docs).
  • Pin runtime and package manager versions (.nvmrc/.tool-versions/engines, lockfile committed).

3. Scaffold

  • Use the framework's official generator when one exists; never hand-copy boilerplate from memory when a generator is available.
  • Repo: git init, .gitignore for the stack, README with run commands, license if public.
  • Env: .env.example with every variable name and a comment, no real values; .env ignored. Secrets go through credential-broker, never into the repo.
  • Quality: formatter + linter + typecheck, one test runner with a first passing test, pre-commit hook optional.
  • CI: one workflow that installs from lockfile and runs lint, typecheck, test, build on every push/PR.
  • Local run: one command to start (dev script, Compose file for services) - details in local-dev-setup.

4. Build in vertical slices

  • Order slices so the riskiest assumption is proven first (auth, payment, data import, the core screen).
  • Each slice goes end to end: data model/migration, server logic, API, UI, test. It is demoable when it ends.
  • Typical order: skeleton layout and routing → core entity CRUD → auth/roles if any → the main user flow → secondary features → polish (empty/loading/error states, responsive, a11y, SEO).
  • Real data or explicit empty states only; never fake numbers in the UI.
  • Per slice, implement inner layer outward with tests (plan-then-code section B); multi-day builds keep a short plan file.
  • Large independent slices can go to sub-agents with isolation (task-delegation); the main agent integrates.

5. Verify

  • Run <LINT_CMD>, <TYPECHECK_CMD>, <TEST_CMD>, production build; quote the output.
  • Start the built app and click/curl through the main flow; for web, check responsive widths (360, 768, 1440) and keyboard navigation.
  • Public pages: run web-quality-audit (SEO, performance, a11y).
  • Apps with accounts or payments: quick audit via security-audit-cycle.
  • Close with definition-of-done.

6. Hand off

  • First deploy and later releases: release-deploy (choose <DEPLOY_TARGET> from the reference).
  • If the app exposes an agent/MCP interface: agent-tool-parity.
  • Record stack decisions, run commands and open items (living-docs).

Fixed rules

  • Ask before choosing anything that costs money (paid services, plans) or locks in a vendor.
  • No secrets, keys or personal data in code, examples, seeds or commits.
  • Licenses: only dependencies whose license fits the project; note copyleft ones.
  • Three failed attempts at the same step: stop and report (definition-of-done).
  • Do not gold-plate: v1 scope is what the user confirmed; list extras as suggestions.

Report to the user

Result first: what runs, how to start it (one command), what each slice delivered, verification evidence, stack table, decisions taken on their behalf, and the next suggested slice or release step.

相關技能