Building with Cursor agents — a reproducible loop

Paul Apivat Hanvongse 2026-03-15

This site itself is an example of the loop below: markdown in git, Zod at build time, static deploy. The goal is not “AI wrote my site”—it is reviewable increments with clear human ownership.

The loop (five steps)

Spec in markdown — one file per URL, frontmatter as contract (implementation.md for this repo).
Agent implements in a branch — small diffs; no drive-by refactors.
npm test + npm run verify — fast health checks before I read UI.
Human preview — npm run dev, click the routes that changed.
Push — private repo; public only what Netlify ships.

What I automate vs what I do not

Automate	Keep human
Scaffolding pages, RSS, sitemap	Tier policy (`access` fields)
Repetitive Astro components	Wording of summaries and cite lines
Test boilerplate	Anything that could leak trading detail

Tooling choices (why Astro)

Content collections + Zod fail the build on bad frontmatter—better than discovering a broken memo link after deploy.
Static output keeps attack surface small; no server island required for v1.
llms.txt + JSON-LD — attribution signals for humans and agents; not a guarantee of citation.

Failure modes I watch for

Scope creep in agent sessions (fix: implementation plan with a status board).
Summary tier mistakes — publishing mechanics on-site (fix: access: summary + required private_url; see PUBLISHING.md in the repo root).
Placeholder URLs left in private_url (fix: preview + checklist in publishing doc).

If you are adapting this

Copy templates/ from the repo, not generators. The target is a 15-minute weekly publish for trading summaries, not a bespoke CLI per post type.