Produces fact-checked, voice-consistent articles. 7 phases with web search integration and file-based handoff between phases.
OVERVIEW
The content pipeline produces articles where every claim has a source and every draft gets adversarial review before publication. The fact-checker is never the same agent that wrote the piece. Each phase reads previous output from disk, not from truncated session history. Voice style is configured with concrete examples, not vague instructions.
ARCHITECTURE
Web search enabled. Gathers sources across the topic, extracts key findings, and produces a structured research document with citations.
Drafts the article from research output. Follows voice style configuration with concrete examples of tone, sentence structure, and personality markers.
A SEPARATE agent from the writer. Verifies every claim against the research sources. Flags unsupported assertions. Cannot approve its own work.
Adversarial review. Checks for backlash risk, overstatement, missing nuance, and potential misinterpretation. Catches what the fact-checker does not look for.
Addresses all findings from fact-check and red-team phases. Rewrites flagged sections while preserving the voice and flow of the original draft.
Voice consistency pass, flow improvements, and readability optimization. Ensures edits from the fix phase did not break the narrative arc.
LinkedIn/social summary with its own red-team review. A separate deliverable that undergoes independent adversarial checks.
DESIGN
The fact-checker is never the same agent that wrote the piece. You cannot trust a writer to verify its own claims -- the same biases that produced the hallucination will miss it on review.
Each phase reads previous output from disk, not from truncated session history. No context window limits. No lost information between phases.
Configured with concrete examples of desired tone, sentence structure, and personality markers. Prevents the pipeline from stripping personality into corporate blandness.
Truth and trust require sequential verification. Parallel-then-merge produces blended mediocrity. Each phase builds on verified output from the previous phase.
RESULTS