I built a writing editor for my personal site this weekend

I’ve been rebuilding my personal site for a couple of weeks. This post is about one specific feature: a writing editor that lets me draft a post, get it reviewed by Claude against my actual code and docs, and publish directly to sajivfrancis.com — all without leaving the browser. Drafts live in Cloudflare KV — DRAFTS_KV, owner-only behind CHAT_TOKEN. Once published, only the final version goes to GitHub; the site builds from there.

Here’s the loop. One Cloudflare Worker — the chat-worker — handles all read/write to DRAFTS_KV, owner-only access enforced by CHAT_TOKEN.

flowchart TD
    A["/admin/write"] -->|Save| B["DRAFTS_KV<br>owner-only via CHAT_TOKEN"]
    B -->|Publish| C["GitHub: src/content/blog/<br>YYYY-MM-DD-slug.mdx"]
    C -->|git push| D["GitHub Pages rebuild"]
    D --> E["sajivfrancis.com/blog/slug"]

The review pass fetches the actual GitHub source for any repo I tag, then asks Claude to ground my prose against that code. If I write “I used Python” but ‘wrangler.toml’ says TypeScript, the suggestion flags it — I fix it manually before publishing.

This is a different way of writing. Review prose and technical claims in the same pass — fix what’s flagged, ground the build details against the actual repo, publish. The corpus grounding goes further: retrieval pulls from docs.sajivfrancis.com and the pgvector store, so I can cross-check a new post against everything I’ve already written and documented.

sequenceDiagram
    actor Owner
    participant Editor as "/admin/write"
    participant Worker as chat-worker
    participant KV as DRAFTS_KV
    participant GH as GitHub
    participant Claude
    participant PG as pgvector store

    Owner->>Editor: Save draft
    Editor->>Worker: POST /draft (CHAT_TOKEN)
    Worker->>KV: Write draft
    Owner->>Editor: Review
    Editor->>Worker: POST /review (CHAT_TOKEN)
    Worker->>GH: Fetch repo source
    Worker->>PG: Retrieve corpus chunks (topK=6)
    Worker->>Claude: Prose + grounding payload
    Claude-->>Worker: Suggestions
    Worker-->>Editor: Display suggestions
    Owner->>Editor: Publish
    Editor->>Worker: POST /publish (CHAT_TOKEN)
    Worker->>GH: Push MDX to src/content/blog/

The retrieval query is constructed from the draft’s title and the first 800 characters of the body — enough signal to surface relevant chunks without overfitting to the opening paragraph. The actual call looks like this:

// pgvector corpus grounding — use draft title + body lead as the
// retrieval query. Owner mode so all visibility scopes are in play.
let corpusChunks = [];
const retrievalQuery = `${draft.frontmatter.title}\n\n${draft.body.slice(0, 800)}`;
try {
  const ctx = await getContext(
    retrievalQuery,
    ['docs', 'writing', 'meta'],   // ← topic filters
    undefined,                       // owner → no visibility filter
    'owner',
    clientId,
    env,
    { topK: 6 }                      // ← matches topK in component table
  );
  corpusChunks = ctx.chunks ?? [];
} catch (e) {
  // Non-fatal — continue with GitHub grounding only
}

The AI layer is only as useful as its grounding — GitHub source, personal docs, the vector store. Strip that away and it is just a grammar checker.