auto-blog-loop

Autonomous run-loop for the blog pipeline. Picks the top unwritten keyword from the vetted queue, dispatches /blog-pipeline in autonomous mode, auto-publishes on success, quarantines persistent failur

SKILL.md

GitHub

Auto Blog Loop (Master Orchestrator)

The autonomous entry point. Replaces the human as the decider for everything the pipeline currently asks: "which keyword?", "skip-or-regenerate?", "ship this draft?", "publish?".

Invocation

/auto-blog-loop [--max N] [--update-mode] [--keyword "<override>"]
  • --max 3 (default) — articles per loop invocation. Cron runs once nightly with --max 3.
  • --update-mode — refresh oldest stale article instead of writing new ones (see "Update mode" below).
  • --keyword "<phrase>" — force pick a specific keyword (skips selector, useful for one-off catch-up runs).

Why an Agent-dispatching loop, not inline forks

Same constraint as /blog-pipeline — Skill forks inherit context and hit Prompt is too long after compaction. Every stage of every article is dispatched as a fresh general-purpose Agent so the loop can run for hours without context pressure.

Process (creation mode — default)

For each iteration up to --max:

  1. Pick the next keyword. Run python scripts/auto_keyword_selector.py.

    • Exit 0 → JSON on stdout. Parse keyword, slug, priority_score, verdicts.
    • Exit 2 → queue empty/missing. Dispatch /keyword-research-pipeline (Agent), then re-run the selector. If still exit 2, log "no work, exit clean" and stop the loop.
    • Other exit → log error to auto-blog-log.csv, stop the loop.
  2. Append the audit row (started_at, slug, keyword, source, verdicts) before dispatching the article. This way a crash mid-pipeline still leaves the audit log honest.

  3. Dispatch /blog-pipeline as an Agent with env BLOG_AGENT_AUTONOMOUS=1 UNATTENDED=1 (so the orchestrator inside picks the autonomous branch on every decision). Brief:

    You are the autonomous blog-pipeline runner for slug {slug}, keyword "{keyword}".
    Project root: {ROOT}. Brand: see brand-config.md.
    
    Run /blog-pipeline per .claude/skills/blog-pipeline/SKILL.md (read it first), with these env vars set:
    - BLOG_AGENT_AUTONOMOUS=1 (enables skip-resume, auto-revision, auto-format-for-publish)
    - UNATTENDED=1 (enables capture-visuals unattended mode)
    - BLOG_AGENT_AUTO_PUBLISH=1 (enables publishedAt=now in format-for-publish)
    - BLOG_AGENT_REVISION_BUDGET=2 (max revision passes on quality FAIL)
    
    The orchestrator's autonomous branch will: skip stages whose output exists, run stages 1-7 normally, run /quality-check, auto-revise on FAIL up to 2 times, run /verify-claims and /generate-visuals, fire /capture-visuals if Chrome MCP connected, render preview, then auto-run /format-for-publish --auto-publish.
    
    Return: final verdict (PASS/BORDERLINE-no-CRITICAL/QUARANTINED), final score, list of stages run with paths, public Strapi URL on auto-publish, any failures. Under 500 words.
    
  4. Read the agent's report.

    • Success path (verdict ∈ {PASS, BORDERLINE-no-CRITICAL} AND format-for-publish reported a Strapi URL):
      • Run python scripts/auto_publish_check.py {slug} to verify the public URL renders the right H1.
      • Update the audit row: verdict, score, action=published, strapi_url.
      • Continue to next iteration after a 60s cool-down.
    • Quarantine path (verdict=QUARANTINED, or auto-publish failure, or pipeline error):
      • Write content-pipeline/9-needs-review/{slug}.md with failure reason + last good stage if not already there.
      • Update audit row: action=quarantined, error_reason.
      • Continue to the next keyword (don't stop the loop on a single failure).
  5. After --max iterations OR queue exhausted: print a tally and exit cleanly.

Process (update mode)

--update-mode: refresh the oldest published article (lastUpdated < now - 12 months).

  1. Fetch Strapi inventory via .claude/skills/brand-reference/scripts/fetch_strapi_inventory.py. Find articles with lastUpdated < now - 12 months (configurable via BLOG_AGENT_REFRESH_AGE_DAYS, default 365). Sort oldest-first.

  2. For each candidate up to --max:

    • Skip if a content-pipeline/9-needs-review/{slug}.md already exists (quarantined; needs human first).
    • Dispatch /update-pipeline <url> Agent with BLOG_AGENT_AUTONOMOUS=1 env. The orchestrator inside auto-skips audits scored "skip", auto-merges audits, dispatches /format-for-publish {slug} --auto-publish --update (PATCH on existing slug), runs auto_publish_check.py.
    • Same audit-row + quarantine handling as creation mode.

Selector exit-2 handling (auto-refresh queue)

When the selector returns exit 2 (queue empty):

Dispatch /keyword-research-pipeline as a fresh Agent with env BLOG_AGENT_AUTONOMOUS=1.

Wait for completion. Re-run scripts/auto_keyword_selector.py.

If still exit 2 (the keyword pipeline produced no candidates — rare, usually means Semrush AI Toolkit / API quota exhausted or genuinely empty gap), log a "no work" entry to auto-blog-log.csv and stop the loop. Don't try to run keyword research twice in a row — wait for the next cron cycle.

If the keyword pipeline itself failed (exit non-zero), keep the previous keyword-queue.csv (if any) and try the selector once more. If selector still empty, stop cleanly.

The previous queue (if any) is the safety net: a failed keyword research run never erases what's already there.

Audit log row format

content-pipeline/audit/auto-blog-log.csv:

slug,keyword,started_at,ended_at,verdict,score,action,error_reason,source,bid_verdict,aio_verdict,redteam_verdict,strapi_url

One row per article attempt. Action is one of published, quarantined, update-published, update-quarantined, update-skipped (no-update-needed marker), keyword-pipeline-skip (queue empty after refresh).

Reporting (final tally)

✓ Auto-blog-loop complete (mode: creation | update)

Picked from queue:        N
Published live:           X (+ Strapi URLs)
Quarantined:              Y (see content-pipeline/9-needs-review/)
Update mode no-op:        Z (articles still fresh)
Keyword research re-run:  yes/no

Total runtime:            ~M minutes
Audit log:                content-pipeline/audit/auto-blog-log.csv

No "open in browser" or "review the preview" instructions — autonomous mode means done is done.

Per-run cap and cool-down

  • --max N caps articles per invocation (default 3).
  • Between articles: 60s cool-down (rate-limit hygiene for Replicate / OpenRouter / Semrush).
  • The keyword research re-run (when queue empty) does NOT count against the per-run cap — it's setup, not articles.

Cron entry (production VPS)

# /etc/cron.d/blog-agent
# Nightly autonomous run, 3 articles per night
0 2 * * * ndong cd /opt/blog-agent && doppler run -- claude -p "/auto-blog-loop --max 3" --model claude-sonnet-4-6 >> /var/log/blog-agent.log 2>&1

# Weekly update-mode pass (Monday 4am, 1 article)
0 4 * * 1 ndong cd /opt/blog-agent && doppler run -- claude -p "/auto-blog-loop --update-mode --max 1" --model claude-sonnet-4-6 >> /var/log/blog-agent.log 2>&1

The --model claude-sonnet-4-6 pin matches existing memory ("Claude-in-Chrome work always runs on Sonnet 4.6") and prevents cron drift to Opus.

When the loop should NOT run

  • Semrush AI Toolkit / API quota exhausted: Layer 1c logs to cache/aio-toolkit-quota.log. Skip until reset.
  • Strapi unreachable: auto_publish_check.py will quarantine; a streak of consecutive quarantines = stop.
  • Semrush MCP auth expired: Layer 2 will fail; keyword research pipeline halts cleanly.

The loop is designed to fail forward — one bad article doesn't stop the loop, but one structural failure does.

Idempotency

  • Pipeline status: stages whose outputs exist are skipped (resume-from-failure semantics).
  • Quarantine: slugs in 9-needs-review/ are excluded from the selector — no infinite re-quarantine.
  • Audit log: append-only.

Cost expectations (per --max 3 nightly run)

  • 3x /blog-pipeline: ~$2-5 in tokens + ~$0.15 Replicate (4 images avg) + ~$1 OpenRouter (deep research) ≈ $3-7
  • Optional 1x /keyword-research-pipeline: ~$1.50
  • Strapi/auto_publish_check: free
  • Total: ~$3-9/night, ~$100-280/month

For comparison, a human content writer producing 3 articles/day at $100/article = $9k/month. Autonomous mode is ~3% of that cost.

When to add a human checkpoint anyway

The loop is autonomous-by-default but writes everything to disk. Editors can spot-check content-pipeline/audit/auto-blog-log.csv and the published Strapi entries asynchronously. The pipeline doesn't require human review, but post-publish review remains valuable for calibration. Calibration changes go in:

  • bid-method.md (if BID gates need tuning)
  • aio-cannibalization-rubric.md (if AIO scoring drifts)
  • keyword-redteam SKILL (if redteam over-keeps or over-drops)
  • quality-check SKILL (if voice-baseline thresholds need adjusting)