Ghost Autoblogger Bot 🤖

Automated French content for a Ghost blog. Two independent agents:

• News agent — finds fresh AI news with hosted web search, writes the article, then derives a Google Discover–optimized title and its image keywords.
• Guide agent — picks an unpublished evergreen SEO topic and writes a long-form guide for Google Search.

Both publish straight to Ghost with a featured Unsplash image, a random staff author, and a local markdown copy in generated_articles/.

Model strategy

Everything runs on a single model, gpt-5.6-luna, with reasoning effort as the only dial. As of August 2026 the whole capability/cost Pareto frontier is held by Luna effort levels — no GPT-5.4 configuration is the best choice at any budget. The gpt-5.6 alias routes to Sol, so the Luna id is spelled out.

Per-task profiles live in src/config.js — one place for model, effort and output caps:

Task	Effort	Why
newsDiscovery	low	Web search injects ~8.6k input tokens per call; keep the model cheap here.
article	high	The product.
articleMeta	high	The Discover title is the #1 CTR lever.
guide	high	Long-form SEO.
guideTopic	low	Pick a topic, avoid duplicates.
imagePick	none	Selection from a supplied list.

Caps are max_output_tokens and must cover reasoning tokens plus the answer; they sit just above the real envelope so runaway responses surface instead of hiding.

Cost discipline

• Prompt caching: every call splits stable content (instructions) from volatile content (input). Caching matches a prefix, so one changed byte at the front discards the rest. No prompt_cache_key is set — it scopes the cache to a namespace that starts empty and lowers the hit rate.
• Cached share is logged per call: [openai:article] in=1841 cached=1536 (83%) out=2210. That ratio moves long before a bill does; a sustained drop means a prefix drifted.
• Merged calls: title + image keywords are one call, guide topic + image keywords are one call. Each merged prompt states explicitly that its deliverables are judged independently — merging can otherwise collapse one answer into the other.
• Dedup lists are never numbered. Positional labels renumber the whole block whenever an entry ages out, rewriting it from its first byte.
• Ghost reads are memoized in-process (authors, tags, guide archive), so --run does not fetch the same thing twice.

Structured outputs (strict JSON schema) are used everywhere except the search-backed discovery call — JSON mode and the hosted search tool are mutually exclusive, so that one recovers JSON with a tolerant parser (src/utils/json.js).

Setup

bun install
cp .env.example .env

Fill .env:

OPENAI_API_KEY=
GHOST_API_URL=
GHOST_ADMIN_API_KEY=
UNSPLASH_ACCESS_KEY=

Usage

bun run news     # one-shot: news article
bun run guide    # one-shot: SEO guide
bun run run      # one-shot: both
bun start        # persistent scheduler (node-cron)

One-shot mode exits non-zero on failure, so system cron can alert on it. Schedules live in src/config.js (NEWS_CRON, GUIDE_CRON).

Quality gates

bun run check

Runs Biome (lint + format), TypeScript in checkJs mode, and the unit tests. Nothing should be committed with a warning.

Structure

src/
├── index.js              # CLI + scheduler
├── config.js             # model, per-task effort/caps, schedules
├── agents/
│   ├── newsAgent.js      # Discover pipeline
│   └── guideAgent.js     # SEO pipeline
├── services/
│   ├── openai.js         # single API entry point, usage logging
│   ├── newsDiscovery.js  # web-search veille
│   ├── ghost.js          # publishing + memoized reads
│   └── unsplash.js       # image search + AI pick
└── utils/                # json, markdown, cache, env

License

MIT — see LICENSE.