Files

Andrew Yim 6d7a6d7a88 feat(reference-curator): Add portable skill suite for reference documentation curation

6 modular skills for curating, processing, and exporting reference docs:
- reference-discovery: Search and validate authoritative sources
- web-crawler-orchestrator: Multi-backend crawling (Firecrawl/Node/aiohttp/Scrapy)
- content-repository: MySQL storage with version tracking
- content-distiller: Summarization and key concept extraction
- quality-reviewer: QA loop with approve/refactor/research routing
- markdown-exporter: Structured output for Claude Projects or fine-tuning

Cross-machine installation support:
- Environment-based config (~/.reference-curator.env)
- Commands tracked in repo, symlinked during install
- install.sh with --minimal, --check, --uninstall modes
- Firecrawl MCP as default (always available)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

2026-01-29 00:20:27 +07:00

3.0 KiB

Raw Blame History

description, argument-hint, allowed-tools

description	argument-hint	allowed-tools
Review distilled content quality. Multi-criteria scoring with decision routing (approve/refactor/deep_research/reject).	<distill-id\|all-pending> [--auto-approve] [--threshold 0.85]	Read, Write, Bash, Glob, Grep

Quality Reviewer

Review distilled content for quality and route decisions.

Arguments

<distill-id|all-pending>: Specific distill ID or review all pending
--auto-approve: Auto-approve scores above threshold
--threshold: Approval threshold (default: 0.85)

Review Criteria

Scoring Dimensions

Criterion	Weight	Checks
Accuracy	25%	Factual correctness, up-to-date info, proper attribution
Completeness	20%	Covers key concepts, includes examples, addresses edge cases
Clarity	20%	Clear structure, concise language, logical flow
Prompt Engineering Quality	25%	Demonstrates techniques, shows before/after, actionable
Usability	10%	Easy to reference, searchable keywords, appropriate length

Score Calculation

score = (
    accuracy * 0.25 +
    completeness * 0.20 +
    clarity * 0.20 +
    prompt_eng_quality * 0.25 +
    usability * 0.10
)

Decision Thresholds

Score	Decision	Action
≥ 0.85	APPROVE	Ready for export
0.60-0.84	REFACTOR	Re-distill with feedback
0.40-0.59	DEEP_RESEARCH	Gather more sources
< 0.40	REJECT	Archive (low quality)

Review Process

1. Load Distilled Content

source ~/.envrc
mysql -u $MYSQL_USER -p"$MYSQL_PASSWORD" reference_library -e \
  "SELECT * FROM distilled_content WHERE distill_id = $ID"

2. Evaluate Each Criterion

Score 0.0 to 1.0 for each dimension.

3. Generate Assessment

{
  "accuracy": 0.90,
  "completeness": 0.85,
  "clarity": 0.95,
  "prompt_engineering_quality": 0.88,
  "usability": 0.82,
  "overall_score": 0.88,
  "decision": "approve",
  "feedback": "Well-structured with clear examples...",
  "refactor_instructions": null
}

4. Log Review

INSERT INTO review_logs
  (distill_id, review_round, reviewer_type, quality_score,
   assessment, decision, feedback, refactor_instructions)
VALUES
  (?, 1, 'claude_review', ?, ?, ?, ?, ?);

5. Update Status

UPDATE distilled_content
SET review_status = 'approved'
WHERE distill_id = ?;

Decision Routing

APPROVE → markdown-exporter Content is ready for export.

REFACTOR → content-distiller Re-distill with specific feedback:

{"refactor_instructions": "Add more code examples for the API authentication section"}

DEEP_RESEARCH → web-crawler Need more sources:

{"research_queries": ["Claude API authentication examples", "Anthropic SDK best practices"]}

REJECT → Archive Mark as rejected, optionally note reason.

Example Usage

/quality-reviewer 15
/quality-reviewer all-pending --auto-approve
/quality-reviewer 42 --threshold 0.80

3.0 KiB Raw Blame History