This is a first-person maintenance log from one week of keeping aogl.cn alive with Cursor and other AI coding agents. It is not a vendor pitch, not a “10x engineer” fantasy, and not a claim that agents replace judgment. Hero art reuses an existing site archive image from the img2threejs demo (original/img2threejs/bmx-endurance.png) because that pipeline is itself AI-assisted work I already shipped on this domain. Dated 2026-09-14.

What the week actually looked like

Most sessions were not greenfield apps. They were surgical: revise an article JSON hub, thicken an English or Chinese fragment, prepend a slug into data/tech-hub-articles.json or data/life-hub-articles.json, strip leftover Google search pills, confirm a page is not sitting in ARTICLE_ARCHIVE_NOINDEX_SLUGS, and only then think about rebuilds. Cursor excelled when the task had a crisp file contract—“edit these fragments, do not run build-site, print counts.” It failed when I asked for vague “make AdSense happier” without naming the policy file.

Commit cadence stayed intentionally boring: small thematic commits when I asked for them, no drive-by reformatting of fifty unrelated hub cards, no force-push theater. Agents draft; I still decide whether a diff belongs in history. Mid-week I refused several “helpful” bulk regenerations of locale trees because half-finished prose would have been overwritten by build-seo-locale-pages.mjs.

Fragment editing is the real craft surface

On aogl.cn, durable article prose lives in data/articles/fragments/*-en.html and *-zh.html, not in the generated /en/articles/ mirrors. Agents that edit the wrong layer create confidence theater: the preview looks fine until the next locale split. My week’s rule of thumb:

1. Change JSON metadata and fragments first.
2. Keep short ja/ko/fr/ru/ar stubs short on purpose.
3. Pills only to app URLs and /articles/index.html—never invent Google search chips.
4. Run cover generation or build-articles only when publishing HTML is the goal.
5. Re-read scripts/seo-index-policy.mjs before claiming a piece is “primary.”

Thick bilingual bodies are where AdSense review weight sits. Asking an agent to “translate everything fully into seven languages” wastes tokens and produces mushy stubs that pretend to be essays. Better: one honest EN ≥900 words, one honest ZH ≥1400 CJK characters, five short locale cards.

The AdSense remediation loop (honest version)

Remediation is a loop, not a button. This week’s loop looked like:

• Identify thin surfaces — ranking archives already listed in ARTICLE_ARCHIVE_NOINDEX; do not promote them as primary demos.
• Ship original thick notes — EatToday / WordPet / Worlds Greatest style first-person pages with real original/ screenshots.
• Keep ads gated — js/adsense.js only requests the publisher script on indexable article/about/changelog URLs.
• Keep the homepage quiet — slim outbound via the earlier adsense-slim-outbound.mjs philosophy; no bookmark walls pretending to be content.
• Re-check after edits — a new slug must not accidentally join the noindex set; robots meta and loader behavior must agree.

Agents are good at searching the repo for ARTICLE_ARCHIVE_NOINDEX, isContentPrimaryPage, and pill markup. They are bad at sensing whether a paragraph is actually first-person evidence or keyword soup. That judgment stayed with me every night.

What I refuse to automate

• Blind full build-site during drafting — too many writers, one overwrite storm.
• Auto-commit of agent diffs without reading the patch — especially anything touching ads.txt, policy sets, or hub membership.
• Inventing screenshot archives — if original/ does not contain the UI, the article does not pretend it does.
• Mass-translating stubs into fake longform — short locales are a feature.
• Letting agents “fix AdSense” by stuffing outbound links — the remediation direction is fewer low-value outs, thicker originals.

Where Cursor earned its keep

Repetitive mechanical work: mirroring JSON title fields across seven languages, checking that hub JSON prepends did not resurrect demoted listicles like apple-iphone or pc-standalone rails, verifying image paths exist before writing heroImageSource, counting words/CJK after drafts. It also helped me cite real script filenames instead of paraphrasing from memory—important when the public essay claims the pipeline is real.

Where it needed adult supervision: tone. First-person product notes drift into brochure voice if you accept the first polished paragraph. I kept asking for blunt boundaries (“this is not dietary advice,” “battle power ≠ English grades,” “no AdSense approval promises”) because reviewers smell hype faster than syntax errors.

Takeaway after seven days

AI coding agents are a sharp tool for a static multilang site with a strict authoring contract. They multiply throughput on fragments and policy searches. They do not own the content class decision—indexable original vs noindex archive—and they should not own the publish button. If you maintain a similar aogl-style stack, give the agent file-level jobs with explicit non-goals, then keep human review on ads policy, hub membership, and whether the screenshots still match the prose.

Practically, my end-of-week checklist is short enough to paste into a Cursor prompt: verify hero paths exist; confirm hub prepends; reject Google pills; keep EN/ZH thick; leave other locales stubby; do not run full site build unless publishing; print word and CJK counts. That checklist beats another abstract manifesto about “AI collaboration.” The week succeeded when the repository gained honest primary editorials—not when the chat transcript looked clever.

Related