Skip to content
SeoFreshUp Docs

Custom skills

A “skill” is a Markdown file (max 32 KB ≈ 8 000 tokens) that adds your domain knowledge on top of the default LLM prompt. The skill is injected at the end of the system prompt, just before the user content.

Why use skills

Without skills, the AI writes generic content that follows SEO best practices but has no personality. With skills, you can teach the AI:

  • Your editorial charter — tone of voice, allowed/forbidden formulas, brand vocabulary
  • Your domain rules — e.g. “we never recommend competitor X”, “always cite source Y”
  • Concrete examples — 1-2 examples of “good” vs “bad” outputs are 10× more effective than abstract rules
  • Hard constraints — e.g. “always use the imperative mood”, “max 1 emoji per H2”

Skill roles

You can configure a different skill per AI role:

RoleSkill keyEffect
Editorial (audit verdict)editorialCustomizes how the AI judges your articles
Risk analysisriskAdjusts how Core Updates risk is evaluated
RewriterewriteMost impactful — controls voice/style of generated articles
Duplicate detectionduplicatesTightens or loosens duplicate matching
Image genimagesCustomizes the image generation prompt

Configuration

In 🌿 SeoFreshUp → Settings → Custom Skills, you’ll see one textarea per role. Paste your Markdown content. Save.

Skill structure example

# Editorial skill — LWS hosting blog


## Tone of voice
- Direct, technical, never marketing-y.
- Use "you" not "we".
- French: vouvoiement always, never tutoiement.


## Forbidden phrases
- "Discover", "Solution", "Optimal" → too marketing.
- "WordPress is the best CMS" → never absolute claims.


## Required style
- Always cite Google docs URL when discussing Core Updates.
- Code examples in <pre><code> blocks.
- Headings end with a question mark when possible.


## Concrete example
GOOD justification: "Article from 2014 mentions PHP 5.4 (EOL 2015).
Topic still searched (~500/mo) but the technical content is dangerous to follow as-is."
BAD justification: "Article needs to be updated to be more relevant in 2026."

Limits

  • Max 32 KB per skill (~ 8 000 tokens). Beyond that, the plugin truncates silently.
  • Skills are injected on every AI call for that role — so a 32 KB skill = +8 000 tokens × N articles. Watch your AI bills.
  • Practical sweet spot: 200-500 words per skill. Less is more.

Don’ts

  • Don’t try to reproduce the default rules (the model + verdict scheme are already enforced)
  • Don’t ask the AI to break the JSON output format — the schema always wins
  • Don’t paste a 50-page style guide. Less is more. 200-400 words of specific instructions beats 5000 words of generic advice
  • Don’t include actual content from your articles in the skill — it’ll bias the AI to copy that style verbatim

Cost impact

A 500-word skill = ~125 tokens. At gpt-4o-mini’s $0.15/1M input rate:

  • Per AI call: +$0.000019 (negligible)
  • Per 1000-article audit: +$0.019 total

So skills add ~2 cents per full audit. Worth it for the quality boost.

Iterating

  1. Write a draft skill (~300 words)
  2. Run a rewrite on a single test article — observe the output
  3. Adjust the skill based on what you don’t like
  4. Repeat until the AI consistently writes “in your voice”

This usually takes 3-5 iterations to dial in.

What’s next?