Skills

A “Skill” represents one discrete prompt-and-extract operation run by an LLM in our system.

Skills Runtime

Skills are invoked dynamically. core/llm.py loads the prompt text directly from skills/<skill_name>/SKILL.md. This allows prompt-tuning to happen independently of Python application logic, inside markdown files that AI assistants (like Copilot) can natively parse format-wise.

Skill Catalog

Content Classification

• Purpose: Assign a general topic bucket for filtering (e.g., tutorial, opinion, news, release).
• Inputs: Candidate text.
• Outputs: A mapped category.
• Prompt Location: skills/content-classification/SKILL.md.
• Failure Mode: Item is categorized as Unknown.

Relevance Scoring

• Purpose: Act as tie-breaker for items sitting in the ambiguous cosine-similarity band (0.50 - 0.85).
• Inputs: The project’s topic description, precomputed reference similarity score, candidate title, candidate text (trimmed to 5000 chars).
• Outputs: A JSON payload containing a relevance_score and explanation.
• Prompt Location: skills/relevance-scoring/SKILL.md.
• Failure Mode: Aborts and relies purely on the Cosine Similarity score.

Deduplication

• Purpose: Determine if a new post is functionally identical (or a direct repost) to one already stored recently.
• Inputs: Candidate text and closest-distance match.
• Outputs: Boolean flag suppressing the clone.
• Prompt Location: skills/deduplication/SKILL.md.

Summarization

• Purpose: Condense an input article into a fast 2-3 sentence overview that editors read on the dashboard.
• Inputs: Candidate text.
• Outputs: Short paragraph saved to Content.summary_text.
• Prompt Location: skills/summarization/SKILL.md.
• Failure Mode: Retries; if fatal, leaves UI text empty requiring manual skim.

Newsletter Email Extraction

• Purpose: Parse raw forwarded HTML emails to decouple multiple article links from the sender’s flavor-text wrapping.
• Inputs: HTML email body.
• Outputs: A list of extracted URLs and parsed titles.
• Prompt Location: skills/newsletter-extraction/SKILL.md.

Entity Extraction

• Purpose: Find proper nouns (people, vendors, technologies) referenced in the item so we can track their authority and mention-velocity.
• Inputs: Candidate text.
• Outputs: Unresolved EntityCandidate names.
• Prompt Location: skills/entity-extraction/SKILL.md.

Theme Detection

• Purpose: Turn clustered articles into human-readable newsletter draft sections.
• Inputs: Groupings of high-velocity related content.
• Outputs: Proposed newsletter headings and a contextual summary for the grouping.
• Prompt Location: skills/theme-detection/SKILL.md.