claude-of-alexandria

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 53 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

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

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

Current Collection

6 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 maintainer-run eval scenarios (12 RED + 12 GREEN) + 6 extended scenarios.

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

12 maintainer-run eval scenarios (6 RED + 6 GREEN) + 8 extended scenarios. 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)

37 maintainer-run eval scenarios (20 RED + 17 GREEN) + 12 extended scenarios (adversarial + stress tests for Philemon, Proverbs, 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

14 maintainer-run eval scenarios (7 RED + 7 GREEN) + 3 extended scenarios. Graduated confidence declared before every answer.

argument-flow

Maps the logical argument of a biblical passage using discourse markers:

• Produces connective-anchored proposition chains
• Calls MCP tools for conjunction and discourse data before composing analysis
• Grounds every interpretive claim in retrieved data

16 maintainer-run eval scenarios (8 RED + 8 GREEN) + 3 extended scenarios. For epistles and discourse-heavy passages.

passage-glossary

Assembles a passage reader with a deduplicated lemma glossary:

• Pulls the lemmas of a passage via morphology, deduplicates them, and renders each distinct headword once
• Grounds every gloss in MCP lexicon data; supports multi-passage spans
• Always English output

Composes existing morphology and lexicon tools into a graded-reader study artifact.

Sub-Agents

Skills delegate specialized work to sub-agents. You do not invoke these directly — skills spawn them automatically.

study-evaluator (Sonnet)
    └── biblical-scholar (Sonnet)
            └── data-retriever (Haiku)

Agent correctness is tested indirectly through skill GREEN suites, plus 11 dedicated RED-phase tests that document bare-model failure modes.

Development Setup

Pre-commit hook

ln -sf ../../scripts/pre-commit.sh .git/hooks/pre-commit

This runs secret scanning, TypeScript typecheck, and server tests before every commit.

Installation

Claude Code (Marketplace)

/plugin marketplace add davebream/claude-of-alexandria
/plugin install claude-of-alexandria@claude-of-alexandria

The MCP server is included and auto-configured.

Claude Code 2.1.217
and later require an explicit opt-in for nested sub-agent delegation. Set the
maximum spawn depth to 3, which supports the deepest current chain
(main → study-evaluator → biblical-scholar → data-retriever).

POSIX shells (bash, zsh):

export CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=3

Fish:

set -x CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH 3

If the variable is unset, scholar agents keep working through their direct-MCP
fallback, but they skip the data-retriever compression step and may report
reduced confidence.

Claude Code (Manual)

git clone https://github.com/davebream/claude-of-alexandria.git
cd claude-of-alexandria
ln -s $(pwd)/plugins/claude-of-alexandria ~/.claude/plugins/claude-of-alexandria

The MCP server runs remotely on Cloudflare Workers.

Claude Desktop

Step 1: Download skill ZIPs from the latest release and add SKILL.md contents as project knowledge.

Step 2: Add to your Claude Desktop MCP configuration:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "claude-of-alexandria-mcp": {
      "command": "npx",
      "args": ["mcp-remote", "https://coa.davebream.com/mcp"]
    }
  }
}

Requires Node.js. Restart Claude Desktop after saving.

Verify Installation

• Claude Code: Run /skills and look for all five skills and /agents for sub-agents
• Claude Desktop: Ask Claude to use query_vocabulary for any biblical book

Reference Server

The MCP server provides linguistic data via Cloudflare Workers + D1 (edge SQLite). No local installation required.

Successful responses include a required provenance object identifying every dataset used on that page. Full license detail is at coa.davebream.com/legal/datasets. The v5 contract also keeps native JSON arrays, explicit modes, and page.next_cursor pagination from v4. See the MCP v5 migration guide.

Skills call these automatically. You can invoke them directly if needed.

Hermeneutical Framework

Skills use historical-grammatical method with explicit theological guardrails:

Acknowledgments

Linguistic foundations: Stephen H. Levinsohn (Greek NT discourse analysis) and the OpenScriptures Hebrew Bible / OSHB (Masoretic Text paragraph data, Westminster Leningrad Codex).

Hermeneutical framework: Historical-grammatical method. The Alexandrian school gave us systematic textual criticism. The Antiochene school insisted interpretation stay anchored to the historical sense. The Reformers inherited both. So do we.

Namesake: Clement of Alexandria (c. 150-215 AD), who demonstrated that rigorous scholarship and faithful theology are not in tension.

Contributing

See CLAUDE.md for development guidelines. The head librarian is strict about TDD.

License

GNU General Public License v3.0

_{6 skills, 6 sub-agents, 136 maintainer-run eval scenarios total — 104 core-CI (53 RED + 50 GREEN + 1 smoke) plus 32 on-demand extended. All 66 biblical books. 358 server unit tests (these run in CI).}