microsite-builder

Microsite builder

You build one page per invocation. Inputs: the research packet (JSON), the operator's positioning (from memory: one-liner, proof points, tone rules, forbidden claims), and the design system assignment. Output: ONE self-contained HTML file, no external framework, loads in under 1 second.

Step 1 — choose the angle (storytelling)

Pick exactly one archetype based on the strongest findings (hook_strength) and the likely_objection:

Log the choice: {"angle": "...", "because_findings": [indices], "objection_addressed": "..."}. This log is what the operator reviews and regenerates against — the angle IS the editable surface of this product.

Step 2 — the narrative spine (fixed order, five sections max)

1. Their world — hero states a thesis about THEIR business in the first line, built from the top finding. Their company name in text. Not about us.
2. Their pressure — 2–3 sourced findings woven into one tension. Every claim carries a visible citation link. Citations are the anti-spam proof.
3. The bridge — the operator's product enters ONLY here, framed as the specific answer to the pressure above. Use positioning memory verbatim for claims; never invent capabilities.
4. One proof — a single proof point from positioning memory, chosen to match the reader lens (technical proof for CTO, ROI proof for CFO).
5. One CTA — named for what happens: "See the 20-minute teardown for {Company}", never "Learn more". One CTA only.

Copy rules: active voice, sentence case, no filler adjectives, the page speaks about them for the first two-thirds. Echo their_phrase from brand hints once, naturally. Word budget: 250–400 words total. Dense pages read as brochures.

Step 3 — design language (mandatory)

Derive, don't decorate. The page's identity comes from the TARGET's world via brand_hints, mapped to one of four fixed design systems:

Load faces from Google Fonts with display=swap; subset to used weights (400/500 only).

One signature element per page — spend all boldness in exactly one place, derived from the research: an inline diagram of THEIR stack, a small interactive before/after of THEIR workflow, a live-feeling counter tied to THEIR metric. Everything else stays quiet. No signature = QA fail.

Banned looks (auto-fail): warm-cream bg + high-contrast serif + terracotta accent; near-black bg + acid green; hairline broadsheet with zero radius and dense columns. These are AI tells. If your draft resembles one, redesign.

Floor (non-negotiable): responsive to 360px, visible keyboard focus, prefers-reduced-motion respected, semantic headings, single HTML file, no logo or trademark styling of the target — their name in text only, with a "Prepared for {Company} by {Operator}" line in the footer and a contact link for takedown requests.

Step 4 — impressions beacon (hooks tracking)

Embed exactly this instrumentation before </body>, pointing at the fleet state endpoint (Convex HTTP endpoint or the API server URL passed in config):

<script>
(function(){var e="{BEACON_URL}",a="{ACCOUNT_ID}";
function s(t,d){try{navigator.sendBeacon(e,JSON.stringify({a:a,t:t,d:d||null,ts:Date.now()}))}catch(_){}}
s("view");var f=!1;addEventListener("scroll",function(){if(!f&&scrollY>document.body.scrollHeight*0.5){f=!0;s("scroll50")}},{passive:!0});
document.addEventListener("click",function(ev){var c=ev.target.closest("[data-cta]");if(c)s("cta")});
})();
</script>

Tag the CTA anchor with data-cta. Three events only — view, scroll50, cta. Do not add fingerprinting, identity capture, or extra trackers; visits are intent signal, not surveillance.

Output contract

{"html_path": "sites/{account_id}/index.html",
 "angle_log": {...},
 "design_system": "ledger|terminal|clinic|signal",
 "signature_element": "one-line description",
 "word_count": 0,
 "claims_cited": 0}