Example Generator — scoping note

Settled before this note

• It is a marketing explainer — illustrative only. No engine, no engagement, no committed seed.
• The Business Analyzer is the template — not the Challenge. We convert a visitor; we do not collect a corpus.

The seven decisions

D1 — Input shape

Fields: What is the work? · Who is it for? · What does "done" look like? (optionally a fourth: Any constraints?), each with a short worked placeholder.

Reasoning. A prospect who fills these in has met the seed's spine before they ever sign up — the input teaches the model the prospect should be holding. A single box is simpler but throws that away. Keep it to three or four fields; a teaser, not an induction.

Override: single free-text box for the lightest surface.

D2 — The cross-domain chip set

The seven: a legal matter, a clinical-trial protocol, a research synthesis, a product specification, a regulatory filing, a documentary, and an architecture decision record (ADR). Each chip pre-fills the D1 fields with a short worked seed.

Reasoning. The chips carry the glossary's anti-narrowing discipline: a chip set all of one kind teaches the prospect Loomworks is for that. The ADR is the deliberate software-domain chip — chosen over API docs, runbook, or design system because those read as "a nicer docs tool" a developer already half-has, whereas the ADR names a pain with no off-the-shelf substitute: why did we decide this? Its seed leads on that hook, showcasing the parts of Loomworks with no competitor — provenance, considerations kept not just verdicts, corrections preserved alongside the decision, relationships between decisions. It is the chip that shows recall, the real product. The documentary stays as the chip proving Loomworks reaches creative/non-technical work.

D3 — The grounding text

The prompt carries the four rooms, the lifecycle vocabulary (engagement, assertion, held/committed, consideration, cadence, provenance), the authority posture (the Operator approves; the system surfaces and signals), and the plain-terms discipline — drawn from the glossary.

Reasoning. FORAY's whole technique is a vocabulary-grounding prompt; we have its content. Embedding rather than fetching the glossary at runtime is simpler with no live dependency; the prompt carries a distilled grounding, revised deliberately when the vocabulary settles.

Note. The glossary earns a second job — human definition and AI grounding text. When its Shaping/Rendering entries leave provisional (batches 7–8), the prompt's room descriptions get a matching pass.

D4 — Output section structure

1. Your work in Loomworks — the engagement framing
2. What its Memory would hold — assertions, provenance, kinds of contribution
3. Organizing and arranging it — Manifestation → Shaping for their readers
4. What you'd produce — Rendering: the artifact or the spec
5. Why this helps — recall, provenance, governed approval, in their terms

Reasoning. Mirrors the Business Analyzer's sectioned-HTML contract, mapped to the four rooms plus a benefits close where the prospect's "why care" is answered in their own domain.

Override: add/drop/rename sections.

D5 — Hosting and wiring

• Home: loomworks-marketing (Cloudflare Pages).
• Shape: a Cloudflare Worker bound to the Pages project, serving the AI endpoints at the same origin as the page. Not the FORAY M4-tunnel pattern.
• Key: a Cloudflare environment secret in the Worker. The API key never reaches the browser — hard constraint.
• Model: confirm the current model string at build, not FORAY's pinned value.

Reasoning. A public marketing surface must be always-on. The FORAY tools run a server on the M4 behind a tunnel — fine for a demo, but it makes the page's headline feature depend on a local machine being awake and connected. The marketing page is Cloudflare Pages (always-on, global); a Worker keeps the generator's uptime identical to the page's, with no machine behind it. It is genuinely same-origin (no proxy hop, no CORS dance), holds the key as an edge secret, scales to zero when idle and to spikes under load, and brings edge rate-limiting and bot protection — exactly what an unauthenticated public AI endpoint spending API credits needs. The one cost is a new deploy shape for CC versus the FORAY Flask server: a one-time setup tax against a permanent always-on win. The M4-tunnel pattern stays right for internal/demo tools — but this is the shop window, and the shop window cannot depend on the laptop being awake.

D6 — Persistence

Output retention is client-side only: copy, download, print-to-PDF.

Reasoning. A sales aid, not a corpus. The protocol's auto-capture pattern exists if harvesting is ever wanted, but that carries privacy weight (retaining descriptions of strangers' businesses) and is a separate, later decision.

Override: opt into harvesting now — but that pulls in a storage decision and a privacy posture, expanding scope.

D7 — One chain step at launch, or both

Reasoning. The two-step chain is the technique that makes the output bespoke rather than templated — the whole reason to copy the Business Analyzer over a one-shot generator. Analyze-only first would launch the weaker half; the second call is little more build (same server, same grounding, a second prompt).

Override: analyze-only at launch, illustration as a fast follow, for the smallest first cut.

Scope summary (if all recommendations stand)

A page on loomworks-marketing with three-to-four seed-spine input fields and six cross-domain example chips; a same-origin server making two chained Anthropic calls (analyze → illustrate) grounded in a glossary-derived system prompt; five-section HTML output rendered in-page with copy/download/print; the API key server-side behind a Tunnel; nothing persisted.

Out of scope (named, so they are not assumed in)

• Any engine call, engagement creation, or seed commit.
• Any persistence, library, or harvesting of visitor input.
• The Operator Layer and the engine repos — marketing-site only.
• The Phase 3 API documentation grind — a separate track.

What the CR will need to resolve

• D5 is settled to a Cloudflare Worker; the CR confirms the Worker-to-Pages binding and deploy steps.
• The exact glossary-derived system-prompt text (drafted in the CR, reviewed against the glossary).
• The six chip seeds' wording (adapted from the glossary examples).
• The current model string.
• Rate-limiting / abuse posture for a public, unauthenticated AI endpoint — this is a public surface spending API credits, so the CR should decide a basic rate limit.