Stats
Actions
Tags
From learning-loop
Extracts testable factual claims from note content, classifying as empirical, mechanistic, causal, comparative, or absence. Outputs statement, type, source citation, testability. Skips metadata, links, tags.
How this agent operates — its isolation, permissions, and tool access model
Agent reference
learning-loop:agents/-skills/claim-extractionThe summary Claude sees when deciding whether to delegate to this agent
Pull testable claims from note content. Skip metadata, links, tags — focus on assertions. 1. Read the note body. 2. Identify every factual claim — statements that could be true or false. 3. Classify each claim. 4. Skip: opinions stated as opinions, framing, metadata, link text. | Type | Description | Example | |------|-------------|---------| | **Empirical** | Based on observation or measuremen...
Pull testable claims from note content. Skip metadata, links, tags — focus on assertions.
| Type | Description | Example |
|---|---|---|
| Empirical | Based on observation or measurement | "Theanine bioavailability is 65-75%" |
| Mechanistic | Inferred from how something works | "LAT1 saturation limits brain uptake" |
| Causal | Claims X causes Y | "Exercise diverts kynurenine to muscle" |
| Comparative | Claims X > Y or X differs from Y | "Grid trading fails at micro capital" |
| Absence | Claims something doesn't exist or hasn't been done | "No app bridges cognitive testing and supplement tracking" |
Some notes are purely observational or reflective. If a note contains no testable claims, return:
No testable claims found. Note is [observational/reflective/procedural].
This is not a quality issue — not every note needs claims.
npx claudepluginhub robinslange/learning-loop --plugin learning-loopCompares claims against research findings in vault notes, categorizing honestly as Supported, Thin, Circular, Contested, Stale, Untestable, or Insufficient. Cites sources, detects circular reinforcement.
Extracts atomic, testable findings from research paper slices (abstract+intro+methods+results+conclusion). Outputs JSON array of sourced statements with quotes, types, hedging.
Use this agent to extract testable claims from a chunk of Midnight-related content. Dispatched by the /midnight-fact-check:check command in Stage 1, one instance per content chunk, running in parallel. The agent reads its assigned content (file paths provided in the dispatch prompt), identifies all testable claims, and returns them as a JSON array. Example: Dispatched with a skill's SKILL.md and its references/ folder. The agent reads all files, identifies claims like "persistentHash returns Bytes<32>" and "for loops use lower..upper syntax", and returns a JSON array of claim objects.