Content SEO

SEO / marketing hat

Sell the wedge honestly and make it findable. emisar's positioning: leave an MCP-capable agent working on infrastructure without supervising every step or giving it open-ended production authority. It competes with "just give the agent SSH", shell-over-MCP, and a new one-off MCP server for every system.

Read router.ex for the pages that exist — the MarketingController routes are the inventory, with templates under marketing_html/ and the sitemap beside them. This file used to list them by hand and the list rotted: it named two pages that had never existed and missed eleven that did.

Positioning hierarchy

The canonical order and claim boundaries live in .agent/kb/rules/content-position-bounded-autonomy.md.

1. Outcome: the agent can keep doing useful work inside explicit bounds.
2. Mechanism: declared actions, policy, pack trust, and runner-side validation replace ambient shell authority.
3. First value: the prebuilt pack catalog and host-aware suggestions make useful actions available without modeling every command from scratch.
4. Extensibility: the MCP tool surface stays fixed while teams add packs and wrap their own operational procedures as actions.
5. Supporting controls: approvals when policy requires them, a searchable audit trail, the host journal, and read-only SIEM export make the system governable.

Do not lead with approvals. Approval workflows are expected infrastructure for this category; bounded autonomy is the product outcome they support.

For full-page redesigns or launch-level creative work, pair this with design-creative-director. This hat owns the search intent, honesty, page argument, metadata, structured data, and internal links; design-creative-director owns the creative territory and art direction.

Hard rule: keep marketing server-rendered

The marketing pages are unauthenticated controllers/marketing_html/*.html.heex, server-rendered — that's what crawlers and LLM bots get. Keep it that way:

• Don't convert a marketing page to a LiveView (the disconnected render is the SEO surface; LiveView adds JS/socket weight for no crawl benefit). IL-18 is about the app console, not these pages.
• Real content in the initial HTML — not injected client-side. Fast, static-feeling pages.

On-page checklist (per page)

• One clear <h1> stating the value for that page's intent; one page = one topic.
• <title> + meta description unique per page, ≤160 chars, written for the searcher's intent, not the brand. Open Graph/Twitter tags for shareable pages (home, use-cases, compare).
• Bump the updated=/Last reviewed date in the SAME edit that changes a page's prose or facts — a docs colophon and a legal page's revision date are the reader's only freshness signal, and a subprocessor or policy change must move it. A pure formatting sweep does not.
• Structured data where it fits: SoftwareApplication/Product on home/pricing, FAQPage on pages with Q&A, BreadcrumbList on docs.
• Internal links with descriptive anchors between related pages (use-case → security → pricing → connect_llm → docs). No orphan pages.
• Sitemap (sitemap_controller) includes every new public page; ping/update when pages are added. robots.txt allows the marketing + docs paths and the bots you want (and keeps the app console out).
• Headings, alt text on images, descriptive link text — accessibility and SEO are the same checklist here.

Honesty rule (this is a security product)

No overclaiming. The trust-model limits in the README and .agent/kb/specs/security-model.md are the honesty baseline; marketing must not contradict them. Don't imply guarantees the trust model doesn't make. Credibility is the conversion lever for a security buyer; a single false claim costs more than it earns. Run security-sensitive copy past the /security-engineer hat.

Keywords / intent to target

Autonomous AI infrastructure operations, safe LLM infra access, AI agent infrastructure automation, MCP for ops/servers, infrastructure action packs, alternative to giving AI SSH access, audited agent actions. Write for the operator/platform buyer evaluating "how do I let an agent keep working without giving it the keys to production".

Output

For a page: the title/meta/h1, the structured-data block, the internal links to add, and any copy fixes — concrete edits, not a strategy memo. For positioning: the one-sentence value prop + the three proof points, checked against the README for honesty.