curl -fsSL https://agents.onecontext.me/skills/site-builder.md -o .claude/skills/site-builder/SKILL.md

Site builder — one page, one gate

A site-only owner is routed here by start_context (its nextAction points at this skill when the tenant's products are [site]). You build one page (a landing page, a pricing or about page) for one tenant, once. There is no routine, no buffer and no cadence: interview → brief → draft → lint → preview → checker → owner → publish. rules and rules-blog apply in full; blog-checker judges the draft in a separate call; the owner approves in Context; you (or the driver session) call publish after that and never before.

When the tenant already has a Site channel parent from blog-agent (lane:site), the page issue is a child of that parent. When the tenant was set up for a site only (products [site]), the issue's parent is whatever the owner's map says — the E10 launch page hangs off its epic ticket — and you say which.

Before the first message

1. Context start_context; Blog MCP usage_guide + get_capabilities — page_upsert, preview_render, check_record, publish must be in tools[] (else rules-blog §8: say which is missing, attach the payload you would have sent, stop before that step).
2. Tenant. tenant_get {slug} from the brand the owner named. Missing → tenant_create {slug, name, tier, products: ["site", …], locales, default_locale, timezone, mount_path, primary_hostname?, context_epic_id?}; a brand_upsert (voice, byline, design tokens, organization_ld) follows the interview. Existing → reuse; a second tenant for the same brand is never created.
3. Where pages resolve. The renderer serves <host><mount>/<slug> (non-default locales under /<locale>/). The tenant root <mount>/ is the hub view — the article list — not a page. A landing page therefore needs a slug (launch, home, pricing); say so before the owner expects /. With primary_hostname set the page is served on that host, otherwise at https://sites.onecontext.me/t/<slug><mount>/<page>.

Interview — one round, grill-me shape

Ask the whole frontier as numbered ❓ Qn blocks with a ➡️ recommendation each; wait; stop when nothing is assumed. Facts (the site's palette, fonts, logo, product facts, existing copy) are fetched, not asked.

• Goal — the one thing the page must make the reader do (sign up · read the guide · install · buy). One goal, one CTA.
• Audience — who lands here and what they already know; the objection the page must answer first.
• Sections — propose from the goal, in this order, and let the owner cut: hero (what it is, in the product's own words) · how it works (3–5 numbered steps) · tiers or plans (a comparisonTable) · proof (only what the brief backs: a named quote, a real number, a screenshot — or nothing) · trust (keys, data, limits) · pricing (the owner's exact words, placeholders allowed: "free while in preview") · CTA.
• Copy rules — the brand persona's adjectives, anti-adjectives, banned phrases and claims policy (rules-blog §9): nothing the product does not do today, no compliance claims, no invented numbers, quotes or customers. Name limits before benefits.
• Design tokens — the tenant's design-tokens.json (from brand_upsert) is the page's look; propose them from the brand's own site. The page must look like the brand, not like the platform's other tenants.
• Slug, locale, hostname — slug (lower-case, short), locale, and whether primary_hostname / domain_connect is wanted now.

Write templates/page-brief.md filled and show it; the owner's "go" on the brief is the Round's exit. It is attached to the issue as docKind: "brief".

The Context issue

One page = one Context issue, title Site: <title> (<locale>), labels channel:site locale:<l> tenant:<slug> kind:new gate:artifact, parent_id = the Site parent when it exists. Create it before page_upsert — the Blog MCP requires context_issue_id on every page. Acceptance criteria (acceptance_criteria): lint clean · preview attached · checker verdict recorded · owner approval · live URL.

Draft — the Blog MCP page shape

page_upsert {tenant_slug, slug, locale, title, description, sections[], seo {slug, answer, keywords}, context_issue_id, dry_run}. Sections: paragraph · heading (2/3) · list (steps: ordered: true) · comparisonTable (tiers) · callout (pricing, trust) · quote with cite/source (proof) · image (screenshot, with alt) · exactly one appCta (text, href, label). Inline HTML only <em>, <strong>, <a href="https://…"> — <code> is rejected; write key names in <em>.

Pages skip the article-only lint (answer-first, FAQ shape, hub, citations, statistics, artifact); html_disallowed and banned_phrase still reject, first_person_heavy still warns. description ≤ 1000 characters carries the meta description; seo.answer is the one-paragraph answer an engine may lift.

Copy: the hero says what the product is in the product's own words, in two short paragraphs. Steps are verbs. The tiers table has one row per tier and says who publishes. Every claim traces to the brief. The CTA sentence names the action and may link the second destination inline — the button stays one.

Upsert — honour the lint

1. page_upsert {…, dry_run: true}. Fix every severity: "error" at its path; judge the warnings. Two bounces at most (three attempts); a third rejection → post_comment {parent_id: <issue>, body: "ESCALATE: page_upsert rejected 3× — <codes>"}, leave the issue in_progress, stop.
2. The real page_upsert. Keep id, preview_url, url; every later revision passes the same id.
3. preview_render {tenant_slug, kind: "page", title, description, sections, seo, design_tokens} → attach_artifact {parent_id: <issue>, filename: "preview-r<n>.html", title: "Rendered preview, round <n>", docKind: "preview", content: <html>}. Open the HTML: the tenant's fonts, background and accent must be the CSS variables you see.
4. Attach the draft: attach_artifact {parent_id: <issue>, filename: "<slug>.<locale>.md", title: "<title> (<locale>)", docKind: "deliverable", content: <markdown rendering with front matter: page id, slug, locale, url, preview_url>}.
5. One machine-readable line on the issue: blog-page: <id> · slug <slug> · locale <locale> · preview <preview_url> · url <url>

Checker — a separate call

Start a fresh session or subagent whose only inputs are the page payload, the brand profile, the brief and "follow blog-checker" — never this context. It runs content_lint, opens the preview, fetches every URL, traces every product claim to the brief, judges voice, claims policy and structure (H2s as a table of contents, one CTA), and returns a pass · bounce · escalate verdict. Then, on the issue:

• attach_artifact {parent_id: <issue>, filename: "checker-verdict-r<n>.md", docKind: "review"} · check_record {tenant_slug, subject_kind: "page", subject_ref: <issue id>, round, verdict, findings, models} · post_comment with verdict, preview URL, check_id, maker and checker models.
• bounce → fix each finding at its path, re-upsert with the same id, re-render, hand over as round + 1. escalate → stop; the owner decides.
• pass → page_set_status {id, status: "in_review"} and the gate.

The gate — owner approves in Context

update_issues {ids: [<issue>], state: "in_review"}
post_comment {parent_id: <issue>, body: "Checker passed (round <n>) — draft, verdict, preview and models attached"}
request_review {parent_id: <issue>, reason: "<title> · site/<locale> · <words> words · preview <url> · will publish at <url> · maker <model> · checker <model>"}

When the page is one slice of a larger map (a launch), the map's owner may raise the gate on the parent ticket instead — then you attach everything and say on the issue that the gate is raised there. Wait on get_changes {cursor, waitMs: 25000}; never assume approval.

Publish — after done, never before

page_set_status {id, status: "approved"} → publish {tenant_slug, page_id, context_issue_id, assert_context_done: true} → url on the issue → complete_issues with workStats. Then verify: curl -sI <url> is 200 and the HTML carries the title. A hosted root that shows "Nothing published yet." is the hub view, not a failure — say where the page lives. domain_connect / domain_status for a custom hostname follow blog-agent chart step 7.

Benchmark — what a good landing page produces

• page_upsert clean on the first real call (warnings: []), no banned_phrase, no first_person_heavy.
• 350–700 words; 4–6 H2s that read as a table of contents; steps ≤ 5.
• One appCta, placed last; one goal the hero states in its first sentence.
• Every product claim traceable to the brief; the checker's facts table has no no.
• Preview HTML < 50 KB, no external scripts, no image above the fold (the renderer's page view is text-first, so LCP is the hero paragraph), tokens visible as CSS variables.
• Rendered live within one minute of publish; curl -sI 200; the Organization JSON-LD from the brand present in the page.

What this skill never does

Publish or set approved before the Context issue is done; run its own checker in the same session; add a second CTA; write a claim the brief does not back; invent proof; echo or store a key value; create a second tenant for a brand; promise / when the root is the hub view.