Article Quality Playbook

Where quality actually comes from

The plugin’s humanization pipeline handles the mechanical side automatically — sentence rhythm, banned AI vocabulary, structural variety, em-dash discipline. What it can’t do is decide who your articles are for or what they know. Those come from four things you control: the model you pick, the persona in your prompts, the facts in your Knowledge Anchors, and a handful of settings. Get those four right and the pipeline does the rest.

Pick the biggest model you can afford

This was the single clearest difference in our audit. Articles from a flagship-tier model read tighter, stay grounded in specifics, and follow nuanced prompt instructions faithfully. The budget “mini” tiers are genuinely good — but under pressure from complex instructions they drift generic and occasionally misread an instruction in ways a bigger model doesn’t.

A sane rule: use a mini-tier model for volume drafts you’ll review by hand, and a full-tier model for anything that publishes without review. With bring-your-own-API-key pricing, the cost difference per article is cents, not dollars — set it under API & Integrations → Model.

Write a reader persona, not an SEO checklist

The best-performing system prompts in our audit describe a specific reader in a specific situation and how to talk to them. Not “write SEO-optimized content” — instead:

You write for [site], a publication for [who exactly].
Your reader is [their situation — time, money, mood, what they
already know]. Address them as "you". Be [tone]. Never claim
personal experience you don't have. Authority comes from
clarity and usefulness, not from stories.

In the Article Prompt, tell the model what the reader needs to walk away with: concrete do-it-this-week actions, real numbers and timeframes over vague suggestions, roughly what each step costs. An article written at a persona beats an article written about a keyword every time.

Two prompt instructions that punch above their weight

1. Ask for failure modes. Add a line like “Where this advice commonly fails, say so and say why.” This one instruction forces the nuance that separates real expertise from tip-shaped filler — the model has to engage with when its own advice breaks down, which is exactly what a human expert does and generic AI content never does.

2. Demand a next action, not a summary. “Close with the single next action to take, not a summary.” Recap conclusions are the most recognizable AI tell there is; a concrete closing action is the most human ending there is. (The pipeline strips “In conclusion” headings either way — this makes the final paragraph genuinely useful instead.)

Fill Knowledge Anchors — it’s the specificity engine

The difference between “check your state’s rules” and “a discrimination charge must be filed with the EEOC within 180 days” is a Knowledge Anchor. The model is forbidden from inventing figures, so without a fact pool it stays qualitative. Ten to thirty verified facts per Writing Style — real prices, deadlines, thresholds, names — turn cautious generalities into the concrete specifics that make an article worth bookmarking. In our audit, the anchor-backed articles were the ones that read least like AI, because they contain things AI couldn’t have made up.

Use one Writing Style per content type — and for money, health, or legal topics, give that style an explicitly careful register: explain how things work, name where a real professional is warranted, never state figures you haven’t verified.

The settings that matter

• Multi-pass writing: on. Outline → expand produces deeper, better-structured articles than one-shot generation.
• Voice pass: on if the budget allows. It’s the difference between “correct” and “written by someone with opinions”. It roughly doubles per-article API cost — still cents with your own key.
• Quality gate minimum words: near your real target, not the 400 default. If your niche needs 1,500-word articles, set 900–1,000 so a thin draft gets retried and demoted to draft instead of published.
• FAQ frequency: “most” or “half”. An FAQ block on literally every post is template furniture; on ~70% it reads editorial.
• Publish mode: Schedule with a daily cap, so a batch drips out over weeks instead of landing as a same-day wall of posts.

The three-article test

Single articles can each look fine while the site develops tells: the same pet phrase in every third post, every article opening the same way, identical link phrasings everywhere. So don’t judge your setup on one article — open three in a row and read them back to back. Repetition you’d never notice in isolation jumps out immediately. The plugin fights this at the corpus level too (it scans your recent posts for phrases that are starting to repeat and steers new articles away from them), but your persona and prompts are upstream of everything — if the three-article test keeps surfacing the same riff, sharpen the persona or vary your title shapes.