The system reporting
on the platform.

The thesis

A daily briefing of what developers are shipping on a platform is exactly the kind of work a developer adoption function does. So FlareCraft does that work — using the platform itself. The act of building this is the demonstration; the artifact persists as a useful tool. The system reporting on the platform is the platform.

Architecture

Two Workers in one repo, sharing eight Cloudflare primitives plus three external services (Beehiiv for audience, Resend for delivery, AnswerOverflow's MCP server for Discord):

Workers

Two of them: flarecraft-pipeline (cron + AI + writes) and flarecraft-site (this SSR site, reads). Same runtime, different roles. One Worker per concern is the cleaner production pattern.

Workflows

The pipeline is a WorkflowEntrypoint. Each step (fetch → classify+persist → finalize → notify) is a checkpointed step.do with its own retry policy. A step.sleep between persist and digest forces a fresh Worker invocation so the digest doesn't share the heavy step's subrequest budget. Durable execution is a primitive of the platform — no Temporal sidecar.

Cron Triggers

One line in wrangler.jsonc ("crons": ["0 12 * * *"]) registers the daily 07:00 CT run. The cron handler instantiates a Workflow and returns; the Workflow runs durably in its own context.

Workers AI — Llama 3.3 70B

Classifies each Hacker News post with structured JSON output (response_format: json_schema). The schema enforces the exact shape downstream storage expects — no parsing brittleness. Quality matters more than latency for a daily-cron workload, so 70B over 8B.

Workers AI — BGE Base (768d)

Generates embeddings of title + one_liner for semantic dedup. Different model, same inference platform, same binding.

Vectorize

Holds the embedding corpus. Each new item queries topK: 1; if cosine similarity exceeds 0.92 against any existing vector, it's treated as a duplicate and skipped. Catches the case where the same launch gets posted multiple times across HN threads — hash-on-URL would let those through; semantic dedup catches them.

D1

Two tables: briefings (one row per pipeline run) and items (classified posts). Pipeline writes; this site reads in the SSR request path. Same database, two Workers, different bindings — capability-based separation.

R2

Archives raw normalized post JSON keyed by date/source/id. Egress-free storage means the entire corpus can be re-classified later when the prompt or model improves, without re-fetching from HN.

Pages / Static Assets

This site is server-rendered on a Worker via @astrojs/cloudflare, so dynamic data from D1 lands in the request path with no rebuild needed. Static assets (favicons, fonts, brand) are served by the same Worker via the ASSETS binding.

Beehiiv (external)

Public subscribe page at flare-craft.beehiiv.com. Audience layer — owns growth, archive, list management. Beehiiv's programmatic publish API is enterprise-only, which surfaced at runtime; the resulting dual-platform architecture is actually more honest than a single-vendor stack.

Resend (external)

Transport layer for the daily digest. Single POST per send, no SDK weight. Sends from notes@flarecraft.dev via verified domain (SPF/DKIM/DMARC auto-written to the Cloudflare zone by Resend's add-domain integration — a great two-platform handshake).

AnswerOverflow (external) — via MCP

Cloudflare's developer community lives in Discord (89,684 members in the official guild). We can't directly index Discord without violating its TOS, so we go through AnswerOverflow — an open-source community indexer that the Cloudflare server has chosen to surface in. AO exposes a public MCP server, so the integration is server-to-server JSON-RPC, no scraping. We run a battery of platform-keyword searches scoped to the Cloudflare guild (serverId: 595317990191398933), filter by recency (7-day window), and feed the resulting Q&A threads through the same classifier as HN and Reddit — with a new q-and-a angle and a structurally-derived resolved boolean from AO's solution field.

Choices I deliberately didn't make

• Cloudflare Agents SDK. Considered and rejected. Agents SDK is shaped for stateful chat/tool-use agents on top of Durable Objects. FlareCraft is a stateless scheduled pipeline — Workflows is the correct abstraction. Picking a primitive because it's trendy is anti-signal.
• Hash-on-URL deduplication. Cheaper, but lets near-duplicates through. Vectorize embedding similarity catches semantic redundancy — same launch posted in three different threads — which is the actual shape of the dedup problem.
• Per-post Workflow steps. Tempting (clean checkpoints!), but each step is a state-machine transition with metadata overhead. For batch processing the inner loop goes inside one step.do; outer steps mark phases.
• dev.to and Lobsters ingestion. Scoped out for now. The schema is source-agnostic; adding a source is a single function in pipeline/src/lib/ returning SourcePost[]. Hacker News, Reddit, and the Cloudflare Discord (via AnswerOverflow's MCP server) are live.
• Browser Rendering. Most HN posts that matter are link posts; title plus excerpt is enough signal. Browser Rendering would add latency and cost without clear recall gains at this scale.
• One Worker doing everything. Astro v6's @astrojs/cloudflare adapter doesn't natively host a Workflow class export. Two-Worker split is the cleaner production pattern anyway.

Honest tradeoffs and known frictions

• The classification prompt is one-shot. A reflection / verifier pass would cut false positives but doubles inference cost.
• The 0.92 dedup threshold is empirical, not validated against held-out data.
• Real DX frictions hit during the build, named in the spirit of credibility:
  • Wrangler v3 (installed by create-cloudflare) is past EOL; first-day DX surfaces a deprecated tool.
  • Astro v6 silently broke Astro.locals.runtime.env in favor of import [object Object] from "cloudflare:workers" — only visible at runtime.
  • Workers AI's response_format: json_schema returns one of three different shapes depending on model; caller has to coerce.
  • Vectorize's returnMetadata takes "none" | "indexed" | "all", not booleans.
  • Workers Free's 50 subrequest/invocation limit pooled across the workflow — fixed with step.sleep + Workers Paid upgrade.
  • Beehiiv's Posts API is enterprise-tier-only; the runtime 403 forced the dual-platform architecture.
  • AnswerOverflow's web pages are protected by Vercel's security challenge — Cloudflare Workers can't solve those without Browser Rendering. The MCP server is the legitimate path and is, refreshingly, the path AO actually wants AI agents to take.

About the builder

Andrew Moore is a Founder, Product Executive, and Builder. Learn more about his current and prior work here.

Live at https://flarecraft.dev · source on GitHub.