メニュー
Review, position, and polish Agent View Layer (AVL) / agent-first web projects: .agent pages, agent.txt manifests, AI-native rendering, MCP/API positioning, and marketing/category narrative.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ArgentAIOS/argentos-core --skill argentos-agent-first-web-avl-reviewコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Run Codex CLI, Claude Code, OpenCode, or Pi Coding Agent via background process for programmatic control.
Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less memory, LoRA/QLoRA optimization
Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less memory, LoRA/QLoRA optimization
SOC 職業分類に基づく
SKILL.md を表示中
| name | argentos-agent-first-web-avl-review |
| description | Review, position, and polish Agent View Layer (AVL) / agent-first web projects: .agent pages, agent.txt manifests, AI-native rendering, MCP/API positioning, and marketing/category narrative. |
| version | 1.0.0 |
| author | ArgentOS |
| license | MIT |
| tags | ["agent-first-web","avl","ai-agents","product-positioning","web-standards","mcp","marketing"] |
| triggers | ["review AVL","agent-first website","AI-first website",".agent endpoint","agent.txt manifest","agents as first-class internet citizens","APIs are going away","MCP is the hands AVL is the eyes"] |
| metadata | {"imported_from":{"original_name":"agent-first-web-avl-review","source":"upstream skill profile"}} |
Use this skill when reviewing or polishing projects that claim to make websites/applications agent-native, especially frontier-infra/avl, .agent endpoints, agent.txt, MCP-adjacent tools, AI-first websites, and agent commerce surfaces.
The user’s thesis:
Preferred framing nuance:
Do not bluntly claim “APIs will disappear” in public-facing copy unless the user specifically wants the provocative version. A more defensible framing is:
AVL = Agent View Layer.
Best category phrase:
Agent-native rendering.
Avoid reducing AVL to “metadata for AI.” Metadata sounds optional, SEO-ish, and decorative. “Rendering target” sounds infrastructural.
Core analogy:
agent.Strong one-liners:
Use these distinctions when explaining AVL:
Important line:
MCP is the hands. AVL is the eyes.
Distinguish AVL copy from AVL implementation.
curl -i https://site/agent.txtcurl -i https://site/.agentcurl -i -H "Accept: text/agent-view" https://site/curl -i https://site/some-route.agentagent.txt manifest, not SPA HTML fallback.agent companion routes for meaningful human pagesContent-Type: text/agent-view; charset=utf-8; version=1 or equivalentAccept: text/agent-viewLink header and/or HTML <link rel="alternate" type="text/agent-view" href="...">@meta, @intent, @state, @actions, @context, @nav.agent, agent.txt, or Accept: text/agent-view return the normal React/SPA HTML shell, the distributed AVL layer is not implemented even if the public copy describes it accurately.Inspect the repo and docs.
frontier-infra/avl, key paths have included:
README.mdpackage.jsonspecs/avl-agent-view-layer.mdspecs/avl-thesis.mdspecs/avl-auth-thesis.mdlib/avl/types.tslib/avl/serialize.tslib/avl/define.tssrc/index.tssrc/next.tsVerify live agent endpoints if available.
curl -s https://site/.agentcurl -s -H "Accept: text/agent-view" https://site/curl -s https://site/agent.txtContent-Type, route, generated timestamp, TTL, auth/session declaration, sections, and action links.Evaluate the AVL document shape.
@meta: version, route, generated timestamp, TTL, auth context@intent: purpose, audience, capability@state: structured backing data, ideally token-efficient@actions: available actions with method/href/input schema@context: narrative explanation / meaning@nav: self/parents/peers/drilldown.agent is not merely a scraped summary of HTML but a producer-side rendering of the same application/page state.Evaluate auth and security.
Connect to stack positioning.
Provide marketing polish.
.agent endpoint, agent.txt, examples, package install, spec, adopter badge.Hero
The shift
Human view vs agent view
/products/kettle next to /products/kettle.agent.How it works
.agent suffixtext/agent-view; version=1agent.txtagent.ts beside page.tsxThe six sections
@meta, @intent, @state, @actions, @context, @navAuth and trust
AVL vs alternatives
Live proof
https://ainode.dev/.agenthttps://ainode.dev/agent.txtDeveloper adoption
.agent route.When the user asks for a review, produce:
Short verdict:
What I checked:
What the idea really is:
Why it matters:
What is strong:
What to be careful about:
Marketing polish:
Next steps:
When implementing AVL on an existing marketing/docs site rather than merely reviewing it:
.agent entries..agent companion, if agent.txt omits generated companions, or if sitemap/prerender coverage drifts.text/html; .agent URLs and Accept: text/agent-view should return the agent representation; agent.txt must not be an SPA fallback.index.html metadata plus route-level SEO components can leave duplicate title/canonical/OpenGraph/Twitter tags; add singleton head dedupe in prerender and server injection paths, then test it mechanically..agent routing, per-route HTML fallbacks, or Accept: text/agent-view handling. Use pnpm start/the production app server for final smoke when those behaviors live in the server.curl for /, representative human pages, representative .agent pages, /agent.txt, local AVL/badge assets, expected content types, required agent sections (@meta, @intent, @actions), and page-specific alternate discovery links..agent / agent.txt if claimed./.agent is a dotfile path; Express/static hosts may ignore or deny it by default. Add an explicit route or enable dotfile serving for that endpoint, and verify it does not fall through to SPA HTML./.agent discovery links into every route. Post-process or generate per-route HTML discovery links and page-specific badges, then verify no non-root page still has href="/.agent".<noscript> or SPA fallback HTML instead of hydrated React. If a canonical human URL appears to return “agent-style” plain text, inspect the raw response and fallback/no-JS content before assuming content negotiation is wrong. Canonical URLs should stay text/html and styled/crawler-readable; .agent and Accept: text/agent-view should be the only agent representations..agent routes may be denied unless explicitly served with dotfiles allowed, and directory redirects such as /about -> /about/ can bypass route-specific fallback logic. Verify with real curl -i headers for canonical routes, .agent routes, agent.txt, and Accept: text/agent-view..agent routes/content negotiation are implemented in the production Node/Express server, final verification must run against that server, not only vite preview.raw.githubusercontent.com dependencies are brittle at runtime and are easy review flags; copy them under public assets and smoke-test the local URL.<link rel="alternate" type="text/agent-view" ...> or with attributes reordered; verifiers and dedupers should detect semantics, not exact string order..agent path logic drifts quickly across verify-agent-views, route parity, crawler HTML, prerender, and server checks.agent.txt is not SPA HTML; and forbidden stale terms/URLs are absent from prerendered output.