VacationDeals.to — Content Refresh Runbook (weekly)

Keeps the existing ~1,235 blog_posts fresh by UPDATING rows in place — never inserting, never deleting, never changing slugs. Sibling of vacdeals-evergreen (which adds +50 new posts); run this on the off-weeks from evergreen (same-week doubles token load). Full design rationale: docs/seo-audit-2026-09-24/07-vacpac-content-refresh-skill.md.

FIRST-RUN SETUP (once — idempotent, safe to re-run)

The last_refreshed_at column drives selection + cooldown. Apply the additive migration on the VPS before the first real run (no-op if already applied):

ssh root@72.60.126.82 'cd /var/www/vacationdeals && set -a && source .env && set +a && psql "$DATABASE_URL" -f scripts/sql/2026-10-15-blog-last-refreshed.sql'

schema.ts already declares the column (blogPosts.lastRefreshedAt). The companion scripts/update-blog-batch-json.ts sets it on every refresh.

GUARDRAILS (inherited from vacdeals-evergreen — non-negotiable)

Agent waves ≤5. One wave of ≤5 agents per run; agents write once and stop (no self-expansion loops — they tripped the session limit before).
Guarded deploys: pnpm build BEFORE pm2 restart; if build fails, do NOT restart (site stays up on the old build).
$25 all-in cap per run (agent tokens + any price-check fetches). Don't exceed without asking.
Re-validate agent output programmatically (word floor, FAQ count, clean title/meta) — never trust "I validated it." The update-blog-batch-json.ts --dry-run is that gate.
Sitemap-cache trap: never rm .next/server/app/sitemap.xml* (500s the route). sitemap.ts has revalidate=3600; a normal guarded rebuild regenerates it.
Known-walled price sources (vegas-timeshare, holiday-inn, hyatt, branson-travel-group): FETCH_BLOCKED on curl/WebFetch is a verification gap, not a data error — use stealth-browser-mcp or fall back to deals.scraped_at freshness; never "correct" a price down to nothing on a block.

THE "NEVER DEGRADE" RULE (this skill's prime directive)

Never shorten a post below the word bar (≥750 words prose; ≥900 incl. FAQs). Shorter output = validation failure, full stop. The script REJECTS a length regression on a non-thin post.
Never break the anti-AI blog style. Humanization (deliberate typos, grammar slips, casual first-person) stays in the BODY ONLY; title/meta/FAQ stay clean. Do NOT "correct" the intentional body slips.
Never republish without diffing — review the --dry-run field-level diff before the real UPDATE.
Never change a slug (URL stability + internal links + no 404s). The script rejects a slug it can't find.
Never delete/expire content (blog posts are refreshed, not retired).
Never restructure a pillar hub (/timeshare-presentation-deals, /all-inclusive-vacation-deals, the 8 topic hubs) — freshen copy only; they are link targets per CLAUDE.md.
Backup before the batch UPDATE (Phase 0) — UPDATE has no skip-existing safety net.

PHASE 0 — Safety backup

UPDATE-in-place has no undo. Take a fresh dump first (crown-jewel = deal_price_history, but this also captures blog_posts):

ssh root@72.60.126.82 "/usr/local/bin/vacdeals-db-backup.sh >> /var/log/vacdeals-dbbackup.log 2>&1 && ls -t /root/db-backups/*.pgdump | head -1"

(If a current dump from a same-day evergreen run exists, that suffices.)

PHASE 1 — Select the weekly batch (default 20 = 5 agents × 4)

Run the priority query, dump to a scratch PSV so the run log shows WHY each post was picked. Exclude anything refreshed in the last 60 days (cooldown). Base query:

SELECT slug, category, created_at::date AS created, last_refreshed_at::date AS last_refreshed,
       length(content) AS content_length,
       round(
         EXTRACT(epoch FROM now() - GREATEST(created_at, COALESCE(last_refreshed_at, created_at)))/86400  -- age (days)
         + CASE WHEN length(content) BETWEEN 3500 AND 4999 THEN 120 ELSE 0 END                            -- thin boost
         + CASE WHEN category IN ('brands','destinations') THEN 60 ELSE 0 END                             -- price-drift risk
         + CASE WHEN title ~* '20(25|26)|summer|winter|spring|fall|thanksgiving|christmas|new year|labor day|memorial|easter|valentine' THEN 90 ELSE 0 END  -- season/year rot
         + CASE WHEN slug ~* 'cheapest|^best-|-best-|-vs-|deals-under-|presentation-deal' THEN 40 ELSE 0 END  -- commercial ROI
       ) AS priority
FROM blog_posts
WHERE is_published = true
  AND (last_refreshed_at IS NULL OR last_refreshed_at < now() - interval '60 days')
ORDER BY priority DESC
LIMIT 30;

From the 30, pick ~20, capping any single category at ~60% of the batch (soft round-robin) so one category isn't starved. Skip pillar-hub + topic-hub slugs (freshen-copy-only; don't batch-restructure them). Dump the chosen set to slug|category|created|last_refreshed|content_length|priority PSV.

For each chosen slug, export the current row (title, meta, category, content, faqs, internalLinks, content_length, created_at, last_refreshed_at) to a per-slug JSON in a scratch dir — this is the agents' input AND the diff baseline.

Also dump live deal context for price re-verification — active deals with brand, destination, price, url — so agents can confirm quoted $ figures against real current deals.

PHASE 2 — Dispatch ≤5 refresh agents (one wave)

Split the ~~20 into ≤5 chunks (~~4 each), one agent per chunk. Each agent refreshes its posts, writes updated JSON once, and stops. Per post:

Freshen dates/seasons — advance year tokens, rotate season/holiday framing to the current/upcoming window, update "as of <month>" lines.
Re-verify every $ price against the provided live deals data (WebFetch / stealth-browser the partner lander when a figure is load-bearing and not in deals). Correct stale/repurposed prices. Apply the CLAUDE.md cautions: a "$X gift card / resort credit / deposit" figure is NOT the package price; drop placeholder original_price (≥20000 or < price).
Expand thin bodies past the bar — ONLY posts flagged thin; add substantive prose + a data table to clear 750+ words (≥900 incl. FAQs). Never pad a healthy long post.
Refresh interlinks — fix stale/404 authored internalLinks; keep them minimal (the render-time resolver in lib/internal-links.ts adds geo/keyword-pillar + sibling links — do NOT hardcode a big related block).
Fix SERP-drifted title/meta — rewrite where intent or a quoted number drifted; keep the "vacation deals" primary-keyword posture. Title + meta + FAQ text must be clean (no deliberate typos) — humanization is BODY-only.
Preserve style/structure — BLUF box, ≥6 FAQs (aim 8–10), first-person voice, real vacpack data.

Agents return per-slug JSON matching the update-blog-batch-json.ts schema (same shape as blog batches) + a short change note (what changed + old→new price diffs).

PHASE 3 — Re-validate programmatically (trust nothing)

Run the validator built into the companion in --dry-run (Phase 4) AND an independent cross-file check: rendered prose ≥750 words / ≥900 incl. FAQs, ≥6 FAQs each (unique, no cross-page dupes), title/meta/FAQ clean of the deliberate body typos (check those fields only — do NOT flag body slips), slug unchanged, no length regression below bar, any changed $ references a real current deal. Bounce failures back to the agent; never accept a shortened post.

PHASE 4 — UPDATE rows in place (dry-run first)

Commit the refreshed JSON to research/content-refresh/<YYYY-MM-DD>/ (git = source of record) and push. Then on the VPS, run the update companion (NOT the insert script — it skips existing slugs):

cd /var/www/vacationdeals && git pull -q origin main
cd /var/www/vacationdeals/apps/scraper && set -a && source /var/www/vacationdeals/.env && set +a
./node_modules/.bin/tsx /var/www/vacationdeals/scripts/update-blog-batch-json.ts /var/www/vacationdeals/research/content-refresh/<YYYY-MM-DD> --dry-run
# review the field-level diff (title/meta/price/word-count deltas), then re-run WITHOUT --dry-run

The script UPDATEs by slug, sets updated_at + last_refreshed_at = now(), refuses to create a missing slug, and rejects a below-bar length regression.

PHASE 5 — Guarded rebuild

cd /var/www/vacationdeals && git pull -q origin main && set -a && source .env && set +a && pnpm build && pm2 restart vacationdeals-web --update-env && pm2 save

Build fails → do NOT restart. Don't touch the compiled sitemap route; ISR regenerates it with fresh lastmod.

PHASE 6 — Verify sample renders

Curl 3–5 refreshed slugs at top-level (https://vacationdeals.to/<slug>): expect 200, FAQPage JSON-LD present, the corrected price/date visible, "Explore More" internal-links block intact. /blog/<slug> 308s to top-level. Within the hour, spot-check that /sitemap.xml <lastmod> advanced for a refreshed slug (only if sitemap.ts derives lastmod from updated_at/last_refreshed_at — see open item below).

FINISH / report

Commit refreshed JSON (research/content-refresh/<date>/); push.
Update docs/NEXT-ENHANCEMENTS.md (dated entry) + relevant memory.
Consolidated summary: batch (N by category + selection reason) · prices re-verified + corrections (old→new) · thin posts expanded (before→after words) · titles/metas rewritten · rows UPDATEd + last_refreshed_at set · build/verify status + sitemap lastmod · cost vs $25 cap · coverage (running total refreshed / 1,235 + weeks to clear the 644-post stale cohort at this cadence).

OPEN ITEMS (owner decisions — see spec §7)

Sitemap lastmod: wire sitemap.ts <lastmod> to GREATEST(updated_at, last_refreshed_at) so refreshes are advertised to crawlers (recommended; small change — otherwise Phase 6's lastmod check is moot).
Title rewrites: default is agents may rewrite titles; if you'd rather approve title changes (they affect established rankings), have agents propose titles and apply only body/price/date.
Cadence: weekly manual /vacpac-content-refresh, alternating weeks with evergreen. 20/week clears the 644 stale cohort in ~8 months.

vacpac-content-refresh

SKILL.md.claude/skills/vacpac-content-refresh/SKILL.md