Credible Writer Core and Style Packs

What It Fixes

AI writing often fails in three practical ways: it sounds templated, it states facts without sources, or it contradicts itself internally. Generic rewrite tools usually only polish wording, so they can miss the real delivery risks in public documents, academic papers, Xiaohongshu posts, brand copy, and text-to-video or text-to-image prompts. credible-writer treats the goal as producing a defensible draft first, then applying genre-appropriate style.

How It Works

The skill uses a core + style pack design. The core is style-agnostic and always active:
- AI-pattern removal: handles 24 recurring patterns plus 2026 dimensions such as filler phrases, formulaic structure, overly smooth rhythm, and vague attribution.
- Hallucination control: applies an evidence check to factual claims, separates facts from judgments, and requires source tier and independent corroboration for core data.
- Consistency checking: covers title-body promises, terminology stability, metaphor direction, evidence alignment, and self-test option polarity to catch contradictions that factual verification alone misses.

Before drafting, it routes to a styles/ pack for the target genre, such as official documents, deep reports, academic papers, Xiaohongshu, brand copy, voiceover scripts, text-to-video, and text-to-image. The workflow loads the style pack, protects numbers / commands / URLs / quotations / names, chooses a minimal, standard, or aggressive pass, then runs evidence labeling, AI-pattern cleaning, fidelity review, and residual audit. Before delivery, consistency checks act as a hard gate: FAIL blocks shipping, while WARN requires human review.

Boundaries and Caveats

It is best for content where sourcing, format, and internal logic matter, especially reports, official writing, academic work, brand copy, and prompt engineering. When reliable sources are missing, it marks ❓ or reports insufficient evidence rather than inventing data. When rewriting existing text, reduce intensity and expand protected fragments if key details are lost. For cross-genre tasks, define the target register clearly so official documents do not become social posts, and Xiaohongshu copy does not become academic prose.