Files

Andrew Yim 0496262cd5 feat(skills): author root SKILL.md for reference-curator suite + 7 sub-skills

Finish the migration for the dirs the bulk pass couldn't auto-handle:

- 90-reference-curator/SKILL.md: hand-authored suite orchestrator (pipeline overview,
  7-stage table, /reference-curator run modes, install) — the single loadable entry.
- 90-reference-curator/0{1..7}-*/SKILL.md: generated from each sub-skill's desktop/SKILL.md.
- scripts/migrate_skill_root.py: generalized discovery to find nested suite sub-skills
  (rglob desktop/code SKILL.md), so the migrator now handles suites too.

81-mac-optimizer, 91-multi-agent-guide, 94-dintel-bootstrap need NO root SKILL.md: they
are Claude Code plugins whose skill correctly lives at skills/<name>/SKILL.md (validated).
Adding a root SKILL.md there would violate plugin structure.

All SKILL.md repo-wide validate: flat-root=65, suite-sub=7, plugin-skills=3, 0 failures.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-28 01:14:46 +09:00

3.4 KiB

Raw Blame History

name, description, version, author, environment

name	description	version	author	environment
90-reference-curator	Full reference-documentation curation pipeline: discover authoritative sources → crawl → store → distill → quality-review (QA loop) → export to markdown / project files / fine-tuning JSONL. A suite of 7 composable sub-skills with a single orchestrator entry. Triggers: reference curator, curate documentation, build reference library, research pipeline, discover sources, crawl docs, distill content, export reference, 레퍼런스 큐레이션, 문서 수집 파이프라인.	1.0	OurDigital	Code

Reference Curator (90)

A modular suite that turns a topic or a set of URLs into a curated reference library. Six stages run as a pipeline with a quality-review QA loop; each stage is also a standalone sub-skill you can run on its own.

Pipeline

discover → crawl → store → distill → review (QA loop) → export
   01        02      03       04        05               06
                         orchestrated by 07

#	Sub-skill	Does	Root directive
01	reference-discovery	Find + credibility-rank authoritative sources	`01-reference-discovery/SKILL.md`
02	web-crawler-orchestrator	Multi-backend crawl (Firecrawl / Node / aiohttp / Scrapy)	`02-web-crawler-orchestrator/SKILL.md`
03	content-repository	MySQL storage with version tracking	`03-content-repository/SKILL.md`
04	content-distiller	Summarize + extract key concepts	`04-content-distiller/SKILL.md`
05	quality-reviewer	QA loop: approve / refactor / re-research routing	`05-quality-reviewer/SKILL.md`
06	markdown-exporter	Export to markdown / project files / fine-tuning JSONL	`06-markdown-exporter/SKILL.md`
07	pipeline-orchestrator	Coordinates all stages + QA loop + state	`07-pipeline-orchestrator/SKILL.md`

How to run

Orchestrated (recommended) — the /reference-curator command (see commands/reference-curator-pipeline.md):

# From a topic (starts at discovery)
/reference-curator "Claude Code best practices" --max-sources 5

# From URLs (skips discovery)
/reference-curator https://docs.anthropic.com/en/docs/prompt-caching

# Auto-approve + fine-tuning output
/reference-curator "MCP servers" --auto-approve --export-format fine_tuning

Input modes: topic (→ discovery), URLs (→ crawl), manifest (→ resume). Key flags: --depth light|standard|deep|full, --output, --max-sources, --max-pages, --auto-approve, --threshold, --max-iterations, --export-format project_files|fine_tuning|jsonl.

Individual sub-skills — each has a slash command in commands/ and its own root SKILL.md; run any stage standalone (e.g. just discovery, or just export).

Install

cd custom-skills/90-reference-curator
./install.sh            # interactive (storage dir, MySQL optional, crawler backend)
./install.sh --minimal  # Firecrawl only, no MySQL
./install.sh --check    # verify

Full guide: USER-GUIDE.md. Changelog: CHANGELOG.md.

Notes

Crawler backends and MySQL are optional — --minimal runs Firecrawl-only with no DB.
The QA loop (stage 05) gates quality before export; --threshold / --max-iterations tune how strict and how many refinement passes per document.
This is a one-shot-per-topic curation workflow, not an always-on service.

3.4 KiB Raw Blame History