LinkedIn AI Content Agent

An autonomous AI agent that generates and publishes LinkedIn content on four constrained topics. All posts are grounded in research from curated RSS feeds and live web search, verified through multi-source corroboration, and subject to automated quality checks before publishing.

Topics

Topic	Focus
AI Benefits	Real, actionable ideas for getting value from AI systems
AI Guardrails	High-level tactics for responsible AI deployment
Cyber Incidents	Analysis of recent breaches and attacks
Cyber Advances	Emerging cybersecurity technologies and tools

Posting Rules

• Minimum 2 posts per week (scheduler checks twice daily)
• Maximum 4 posts in any rolling 10-day window
• Minimum 72 hours between consecutive posts (configurable)
• Weighted topic rotation to ensure balanced coverage
• AI quality scoring with auto-retry for low-quality drafts
• Posts are blocked entirely when fewer than 2 independent verified sources exist

Research & Fact-Checking

Every post goes through a multi-stage research pipeline before content is generated. The goal is 100% credibility — no hallucinated statistics, no fabricated sources, no unverified claims.

Pipeline (3 Anthropic API calls per cycle, ~4.5 minutes)

1. RSS Feed Monitor — 15 curated feeds polled hourly, articles stored in SQLite. Feeds are assigned trust tiers (authoritative, primary, secondary) that affect corroboration scoring. No API call — this runs on a background cron job.

2. Web Search (API call #1) — Claude searches for current information on the topic angle using the web_search tool. Returns 5–15 specific, source-attributed claims.

3. Rate Limit Cooldown (65 seconds) — Waits for the Anthropic token-per-minute window to reset.

4. Corroboration (API call #2) — A dedicated fact-checking call that cross-references all claims from RSS and web search. Only claims appearing in 2+ independent sources with sufficient trust weight pass through. Statistics and percentages are rejected unless verified by multiple sources.

5. Rate Limit Cooldown (65 seconds) — Second wait before content generation.

6. Content Generation (API call #3) — Claude writes the post using ONLY verified facts from the corroboration step. The prompt explicitly forbids inventing, embellishing, or including uncorroborated claims. Source attribution is required in the post body.

7. Quality Check — Scores the post on 7 criteria including source_grounding and factual_caution. Posts that score below threshold on either are flagged for rejection.

Source Trust Tiers

Tier	Weight	Examples
Authoritative	3	CISA, NIST, FBI, NSA, MSRC, Google Project Zero
Primary	2	Krebs on Security, BleepingComputer, Dark Reading, Schneier, The Record
Secondary	1	Hacker News, TechCrunch

Minimum trust weight for corroboration: 3 (e.g., 2 primary sources, or 1 authoritative + 1 secondary).

Source Blocking

If the research pipeline cannot find at least 2 independent sources that corroborate a claim relevant to the chosen topic and angle, the post is blocked entirely. The dashboard shows a clear message explaining why, and the scheduler will try a different topic/angle on the next cycle.

Attestation

Every published post includes this line at the end:

Sources verified through multi-source corroboration. Full source list available upon request.

RSS Feeds (15 curated sources)

Configured in src/config/feeds.js:

• Cybersecurity: CISA Alerts, Krebs on Security, BleepingComputer, Dark Reading, The Record, The Hacker News, Securelist, Schneier on Security, MSRC, Google Security Blog
• AI: Google AI Blog, OpenAI Blog, MIT Technology Review, TechCrunch AI, Simon Willison