claude-of-alexandria

Claude of Alexandria

AI agent skills for rigorous biblical study, built on tested exegetical principles.

---

Structured frameworks that prevent AI agents from committing exegetical malpractice. Every skill is built with Test-Driven Development: document the failure, build the fix, verify it works.

The Problem

Frontier models make predictable errors when handling Scripture. These are documented by 40 RED-phase tests that run the same prompts without skills and record what goes wrong:

• Fabricating linguistic data from training memory — inventing morphological parsings, frequency counts, and hapax claims without querying actual data
• Inventing arbitrary divisions to satisfy session counts ("8 weeks on Philemon") without checking manuscript markers
• Presenting single frameworks for contested books as if scholarly consensus exists
• No confidence tiering — treating training-data guesses and parser-verified data with equal certainty
• Moralistic drift — "try harder" applications and therapeutic framing without gospel grounding
• Yielding to user pressure — skipping data verification when asked to "just be brief" or "skip the Greek"
• Accepting truncated pericopes — validating famous verses (John 3:16) as standalone units based on familiarity, not discourse evidence
• Genre-blind analysis — applying epistolary methods to wisdom literature, forcing narrative arcs on proverbial collections
• Ignoring ancient manuscript markers like Masoretic paragraph divisions and Levinsohn discourse features
• Auto-selecting options instead of presenting scholarly alternatives with evidence

The Evidence

89 automated tests verify that skills prevent documented failures. Tests run against claude-agent-sdk with live MCP data — not mocked responses.

Phase	Tests	What it does
RED	41	Runs prompts against a bare model (no skills, no MCP, /tmp isolation). Documents what goes wrong.
GREEN	47	Runs the same prompts with skills and MCP enabled. Proves the skill corrects each failure.
Smoke	1	Verifies the skill-to-agent pipeline works end-to-end.

GREEN assertions use an Opus grader for LLM-rubric evaluation plus structural checks (icontains, section presence). If a skill cannot demonstrate that it prevents a documented failure, it does not ship.

Current Collection

5 skills + 6 sub-agents, all production. Coverage: all 66 canonical books.

Skills

biblical-segmentation

Divides biblical books into coherent teaching units with integrity safeguards:

• Refuses impossible divisions (you cannot divide Philemon into 12 sessions)
• Presents multiple scholarly-grounded options
• Validates against Masoretic paragraph markers and Levinsohn discourse features
• Handles contested books with multiple frameworks

24 automated tests (10 RED + 14 GREEN).

pericope-delimitation

Validates whether a proposed passage holds together as a discourse unit:

• Checks boundaries against Levinsohn discourse features (NT) and Masoretic markers (OT)
• Returns verdict: VALID, EXTEND, CONTRACT, or ADJUST
• Recommends the smallest coherent unit if passage is too short

11 automated tests (4 RED + 7 GREEN). Resists memory-based validation of famous passages.

exegetical-notes

Produces exegetical notes for sermon or teaching preparation:

• 10-section schema from literary context through verification
• Parser-verified lexical data (not training memory guesses)
• 4-tier interpretive labels: linguistic, discourse, scholarly, agent assessment
• Genre-graduated redemptive-historical connections (epistles vs. wisdom literature vs. short letters)

20 automated tests (6 RED + 14 GREEN), including stress tests for Philemon, Proverbs, and 3 John.

consult-biblical-scholar

Scholarly Q&A for biblical texts. Three auto-detected modes:

• MEANING — lexical and linguistic explanation
• VALIDATE — checks analogies, illustrations, or claims against text; returns formal verdict
• CROSS-REFERENCE — finds related passages with scholarly evidence

11 automated tests (5 RED + 6 GREEN). Graduated confidence declared before every answer.