research-extraction

Research Extraction

Apply kb-conventions skill for annotation markers, KB directory layout, frontmatter schema, and synthesis impact rules.

Multi-axis, confidence-annotated extraction from a research corpus using a Workflow that fans out extractor subagents and gates progression through a reviewer agent after each axis.

Prerequisites

Add to .claude/settings.json (or .claude/settings.local.json) before running:

{
  "permissions": {
    "allow": [
      "Bash(xberg *)",
      "Bash(uvx marker-pdf *)",
      "Bash(curl *)",
      "Bash(mkdir *)",
      "Bash(cp *)",
      "Bash(cat *)"
    ]
  }
}

For fully unattended runs (CI, cron), launch the Claude Code session with --dangerously-skip-permissions instead.

When to use

• Multiple raw source files (theses, papers, reports) covering one system
• Claims must be traceable to exact source sections with confidence markers
• Extraction spans multiple axes too large for one context window
• You want reviewer-gated, disk-persisted progression (not manual session seeds)
• Resuming a partial run — pass resumeFromRunId (preferred over manual restart; completed phases return from cache instantly)

Core architecture

Workflow
  phase("Step 0")
    → extractor agent: reads ALL sources, writes reading summary + working memory dump
    → reviewer agent: checks summary completeness, returns confirmed facts JSON

  phase("Axis N")   [sequential, one per axis]
    → extractor agent: reads seed (confirmed facts from prior phases), writes extract-<axis>.md
    → reviewer agent: runs checklist, returns {pass: bool, issues: string[], confirmed_facts: ...}
    → if !pass: re-run extractor with issues until pass (max 2 retries)

  phase("Formalization")
    → extractor agent: reads all confirmed extract-*.md, produces code sketches
    → reviewer agent: checks sketch coverage

Each phase writes its output to disk before the next begins. Reviewer is a separate agent — independent read of the output file, no shared context with extractor.