METASCI
A personal Claude Code skill collection in two domains: academic research (paper analysis and writing) and dialogue personas (guided explanation and creative direction). Formerly Meta_researcher.
Structure
Skills in use are organized into five packs under skills/:
skills/ 34 skills
├── persona/ virgil · beatrice · socrates · picasso · feynman-digest
├── metasci_writing/ meta-writing · meta-writing-insight · meta-writing-mapping · meta-writing-blog ·
│ Meta_Introduction · meta-writing-korean · meta-rewriting-korean · meta-proofreading-korean ·
│ meta-rewriting · meta-rewriting-antiai · meta-rewriting-loop · meta-review · meta-review-council ·
│ meta-proofreading · meta-proofreading-evidence · meta-proofreading-phrasebank
├── metasci_extraction/ extraction-knowledge · extraction-logic · extraction-vocab ·
│ extraction-style · meta-styling
├── metasci_slide/ meta-slide-content · meta-slide-design
└── metasci_workspace/ init-research-principles · init-writing-workspace · init-communication-rules ·
init-figure-rules · init-review-lessons
codex-skills/ generated Codex build of the five packs above
legacy-skills/ superseded — kept for reference, not loaded by the plugin
skills/ is the only hand-written copy — the source Claude Code reads.
codex-skills/ is the same skills rewritten for Codex, generated, never
hand-edited:
node tools/build-agent-skills.mjs rebuild codex-skills/
node tools/build-agent-skills.mjs --check fail if the build output is stale
Edit a skill in skills/, rebuild, commit both. Runtime differences the
build cannot express go in agent-overrides/<profile>/.
Installing on a machine
git clone https://github.com/kkhgeo/METASCI ~/.metasci/METASCI
node ~/.metasci/METASCI/tools/install-skills.mjs --apply
Each runtime is pointed at its own build:
| runtime | gets | how |
|---|
| Claude Code | skills/ | symlinks in ~/.claude/skills/ |
| Codex | codex-skills/ | symlinks in ~/.codex/skills/ |
They are symlinks, so git pull is the whole update path. Add --replace if
real directories are shadowing the links (it backs them up first).
legacy-skills/ holds the earlier writing/ and research/ packs and the retired Hermes build. Where a skill
was carried forward, the metasci_* version is the current one; see
legacy-skills/README.md.
The detailed sections below document the academic-research skills; the persona skills are summarized under "Persona skills".
Overview
┌──────────────────────────────────────────────────────────────────────────┐
│ │
│ Research Paper (PDF) │
│ │ │
│ ├──→ extraction-knowledge → Knowledge_{topic}/ │
│ │ (cited knowledge claims) │
│ │ │
│ ├──→ extraction-logic → Logic_{topic}/ │
│ │ (structure, argument flow, sentence frames) │
│ │ │
│ ├──→ extraction-vocab → Vocab_{topic}/ │
│ │ (POS word inventory, technical glossary) │
│ │ │
│ ├──→ extraction-style → <corpus>/papers/<slug>/ │
│ │ (frames, style vocabulary, Style Card) │
│ │ │
│ └──→ meta-writing │
│ (academic section writing) │
│ ↓ │
│ meta-review │
│ (multi-reviewer draft improvement) │
│ meta-rewriting │
│ (one-shot style transfer) │
│ │
│ Dataset + Literature │
│ └──→ agentic-research │
│ (autonomous data-driven scientific discovery) │
│ │
│ Naming Convention: │
│ extraction-{target} → PDF에서 추출 (knowledge, vocab, logic) │
│ meta-{action} → 글쓰기/스타일 (writing, rewriting, styling, │
│ review) │
│ │
│ Analysis Layer Separation: │
│ extraction-vocab → WHAT words are used │
│ extraction-style → HOW the writing is built │
│ extraction-logic → HOW arguments are structured │
│ extraction-knowledge → WHAT knowledge is cited │
│ │
└──────────────────────────────────────────────────────────────────────────┘
Installation
As a plugin
This repository is a Claude Code plugin (metasci); skills load from the category folders under skills/ (configured in .claude-plugin/plugin.json). Point Claude Code at a clone with claude --plugin-dir /path/to/METASCI.
Manual copy
Or copy individual skill folders (e.g. skills/persona/virgil/) into ~/.claude/skills/ for global access, or a project's .claude/skills/. The trees below illustrate per-skill file layouts:
~/.claude/skills/ # Global (available in all projects)
├── extraction-knowledge/
│ ├── SKILL.md
│ └── references/extraction_template.md
├── extraction-vocab/
│ ├── SKILL.md
│ └── references/extraction_template.md
├── extraction-logic/
│ ├── SKILL.md
│ └── references/extraction_template.md
├── meta-writing/
│ ├── SKILL.md
│ ├── writing.local.template.md
│ └── references/{writing_template,section_guides,citation-and-verification}.md
├── meta-rewriting/
│ ├── SKILL.md
│ └── references/{blueprint-template,journal-styles,output-formats}.md
├── extraction-style/
│ ├── SKILL.md
│ ├── references/{stage0-prep,lens-architecture,lens-vocabulary,lens-card,
│ │ manifest,aggregate,lens-korean}.md
│ └── scripts/{prep,wordlist,quant_check}.py
├── meta-styling/
│ ├── SKILL.md
│ ├── references/revision_guide.md
│ └── scripts/quant_check.py
├── meta-review/
│ ├── SKILL.md
│ └── references/review_template.md
└── agentic-research/
├── SKILL.md
├── README.md
├── references/{world_model_spec,report_template,domain_geoscience}.md
├── scripts/{orchestrator,profile_dataset,init_world_model,...}.py
└── templates/world_model_template.json
Skills
Extraction — read papers, produce reusable material
| Skill | Description | Output |
|---|
| extraction-knowledge | Cited knowledge into 5 epistemological categories | Knowledge_{topic}/ |
| extraction-vocab | Exhaustive POS word extraction + technical term glossary | Vocab_{topic}/ |
| extraction-logic | Structure, argument logic, sentence frames | Logic_{topic}/ |
| extraction-style | Reference papers → style corpus: sentence frames, measured vocabulary, Style Card | <corpus>/papers/<slug>/ |
Writing — produce prose that does not exist yet
| Skill | Description | Output |
|---|
| meta-writing-mapping | Settle section/paragraph composition before any prose | outline.md |
| meta-writing | Draft a section from own figures/tables/data + literature | English + Korean draft |
| meta-writing-insight | Interview the author, author picks the argument frame, then draft (or rebuild an existing draft) with meta-writing | English + Korean draft, insight/ |
| meta-writing-blog | Paper or manuscript → KEI Korean HTML brief | portable folder |
| Meta_Introduction | Environmental-science Introduction blueprint → draft → CARS evaluation | Introduction package |
| meta-slide-content | Source material → approved slide content specification | content spec |
| meta-slide-design | Approved spec → Korean HTML deck in the Onto house style | HTML + PDF |
Revision — work on text that already exists
| Skill | Description | Needs |
|---|
| meta-rewriting | One paragraph: diagnose + 3 complete alternatives | nothing |
| meta-review | Draft vs reference extractions, per-paper reviewer agents | _logic.md / _vocab.md |
| meta-review-council | Claude, Codex, and Gemini review the reasoning independently and rank each other anonymously; the session chairs | logged-in claude, codex, agy CLIs |
| meta-rewriting-loop | Monte Carlo optimisation over N references | N references |
| meta-proofreading | Reviewer panel + judge; whole draft, section, or paragraph | invoked by name |
| meta-proofreading-evidence | Are these phrases actually used in published papers? | nothing |
| meta-proofreading-phrasebank | One English paragraph vs the built-in Academic Phrasebank, by rhetorical function | nothing |
| meta-proofreading-korean | Korean academic/policy draft → rubric-tiered panel proofreading | Korean draft |
| meta-styling | Apply a style corpus to a draft: structural diff, then candidate revisions | a style corpus (N≥1) |
| meta-rewriting-antiai | Strip AI writing tells across 4 layers | nothing |
| meta-mywriting-korean | Korean draft → the author's own style Blueprint | Blueprint |
Workspace — set up a manuscript folder
| Skill | Installs | Read |
|---|
| init-research-principles | first block of CLAUDE.md + AGENTS.md: assumptions first, simplest sufficient approach, surgical changes, verifiable completion, author and file rules | always |
| init-writing-workspace | second block of CLAUDE.md + AGENTS.md: scholarly-reviewer identity, ledger-based completeness, _review/ discipline (installs the principles block too) | always |
| init-communication-rules | AGENT_communication.md: chatbot-tell removal for replies to the author | session start |
| init-figure-rules | AGENT_figures.md + figure_spec.yaml + figspec.py: figures drawn from a pt/mm spec and audited against it | before figure work |
| init-review-lessons | AGENT_review_lessons.md + review_lessons/: one lesson line per outside-AI review finding, one file per section | when a review comes in |
File naming inside a workspace: CLAUDE.md and AGENTS.md are the fixed
entry points the runtimes look for. Every other file the agent reads as
instructions carries the AGENT_ prefix followed by lowercase words joined
with _ (AGENT_communication.md, AGENT_figures.md,
AGENT_review_lessons.md). Data and code keep plain names
(figure_spec.yaml, figspec.py, review_lessons/).
CLAUDE.md is built from marker-delimited blocks (<!-- metasci:<name> vN -->
… <!-- /metasci:<name> -->), one per skill. Each skill touches only its own
block, so the order in which the skills are run does not matter.
Each skill's description names where to send work that belongs to a different
one, so routing between them is in the skills themselves, not only here.
The numbered sections below predate several of these and cover a subset — the
skill's own SKILL.md is authoritative where they disagree.
Persona skills
| Skill | Description |
|---|
| virgil | Adaptive incremental explanation — guides through a topic one segment at a time, pausing for confirmation |
| beatrice | One-shot complete explanation — the whole logical arc at once; sister skill to virgil |
| socrates | Socratic dialogue through maieutic questioning |
| picasso | Visual director — turns concepts into image-generation prompts |
| feynman-digest | Teach-Back digest — restates the prior explanation from the listener's perspective |
1. extraction-knowledge
Extract core knowledge claims from research paper PDFs, classified into 5 epistemological categories.
| Category | Description |
|---|
| Theoretical Foundations | Core theories, conceptual frameworks, hypotheses, models |
| Empirical Precedents | Data, measurements, experimental results from prior studies |
| Methodological Heritage | Research methods, analytical techniques, measurement tools |
| Contextual Knowledge | Geographic, temporal, policy, social context |
| Critical Discourse | Academic debates, limitations, unresolved issues |
> "Read Chen2024.pdf and save to Knowledge_isotopes"
> "Process all PDFs in papers folder to Knowledge_environmental"
2. extraction-vocab
Exhaustively extract every content word from a paper, organized by section and POS.
- POS Tables: Verb, Noun, Adjective, Adverb — per section with frequency and context
- Technical Glossary: Domain, Methodological, Statistical, Chemical, Instrument, Taxonomic terms
- Multi-word terms: Kept as units (not split into individual words)
- Cross-section analysis: Frequency matrix, section-exclusive words, technical term density
> "Extract vocabulary from Weber2021.pdf to Vocab_geochemistry"
> "Extract technical terms from this paper"
3. extraction-logic
Extract three layers of analysis from academic papers:
- Structure Mapping: Sections → Subsections → Paragraphs hierarchy
- Logic Extraction: Inter-paragraph argument flow + Intra-paragraph sentence logic chains
- Sentence Frame Extraction: Exhaustive rhetorical templates (55+ reference frame types, open taxonomy)
> "Analyze logic structure of Weber2021.pdf"
> "Extract sentence frames from this paper"
Output: Structure tree, logic flow diagrams, sentence frame catalog with [P#-S#] traceability.
4. meta-writing (v0.2.0)
Multi-source academic section writing with clear separation between user's own data and prior research.
Two-tier source model:
| Tier | Type | Role in Paper | Citation |
|---|
| My Data | Figures, Tables, CSV | Description & interpretation target | (Figure 1), (Table 2) — no author citation |
| Knowledge Sources | Knowledge MD, PDF, Web | Comparison & evidence | (Author, Year) with source tags |
Knowledge Sources priority:
| Priority | Source | Tag |
|---|
| 1st | Knowledge folder (markdown) | (Author, Year) |
| 2nd | PDF folder (direct reading) | (Author, Year)* |
| 3rd | Web search (gap filling) | (Author, Year)† |
5-Loop process: Source Scan → Knowledge Reading → My Data Analysis + Gap Check → Synthesis & Writing → Verification
Project config: writing.local.md for per-project paths, figure/table mapping, and writing settings.
> "Write Results section based on my figures and Knowledge folder"
> "이 그림 기반으로 Discussion 써줘"
> "Write Introduction from Knowledge_environmental"
Output: English + Korean dual output, APA 7 references (source-tagged), 4-step verification report.
5. meta-rewriting
One-shot style transfer pipeline: reverse-engineer writing style from a reference paper, generate a Style Blueprint, then apply it to the user's draft in a single session.
- Style Blueprint: 6-dimension analysis (Tone, Sentence Architecture, Logical Flow, Transitions, Vocabulary, Citation Style)
- Gap Analysis: Score each dimension 1-10, identify weaknesses
- Two modes: [A] Sentence-level feedback / [B] Full rewrite
> "이 논문처럼 써줘"
> "Nature Geoscience 스타일로 맞춰줘"
> "Rewrite my draft in the style of this paper"
Output: Rewrite_{topic}/ folder with blueprint, gap analysis, and rewritten draft.
6. extraction-style + meta-styling
A producer and a consumer, not two modes of one skill. extraction-style builds
the corpus; meta-styling applies it. They were split so that the rule
"numbers diagnose, they do not set targets" holds on both sides.
extraction-style: reference paper → one folder per paper holding normalized
text, per-section splits, an exhaustive sentence-frame catalog, a measured style
vocabulary, a lean Style Card, and a manifest.meta-styling: tags the draft's own structure blind, diffs it against the
reference's paragraph and sentence-role map across eight dimensions, then offers
2–3 candidate revisions with a recommendation. Measured rates are a closing
sanity check, never an edit target.
Confidence scales with how many papers the corpus holds: single-source (N=1),
provisional (N=2), corpus (N≥3). At N=1 only absences are enforced.
> "kkh_nitrate_iso.pdf 스타일 추출해줘"
> "이 Discussion 초고를 스타일 코퍼스에 맞춰줘"
7. meta-review
Multi-reviewer draft improvement using pre-extracted logic and vocabulary analysis files.
- Per-paper Reviewers: One subagent per reference paper, running in parallel
- Four Academic Principles: Argument Architecture, Prose Rhythm, Cohesion & Coherence, Academic Register
- Three Intensity Levels: Light (flag only) / Standard (flag + rewrites) / Deep (full rewrite)
- Three Severity Levels: Critical / Major / Minor
> "Review this Introduction paragraph"
> "Deep rewrite this Methods section using all reference papers"
> "Check logic flow only. Light mode."
Output: Review_{timestamp}/ folder with individual reviewer reports, cross-reviewer synthesis, and improved draft with change log.
8. agentic-research
Kosmos-inspired autonomous research framework for iterative data-driven scientific discovery.
- Multi-cycle analysis: Hypothesis generation → code execution → literature grounding → refinement
- World Model: Maintains evolving JSON state of discoveries, hypotheses, and evidence
- Literature integration: Combines data analysis with web search for scientific context
- Python scripts: Dataset profiling, world model management, report generation
> "Analyze this dataset and find interesting patterns"
> "Run agentic research on my environmental data"
> "Iterative hypothesis testing on this CSV"
Output: Research_{topic}/ folder with world model, analysis notebooks, and structured report.
Typical Workflows
Full paper analysis
1. extraction-knowledge → extract cited knowledge
2. extraction-vocab → build word inventory + technical glossary
3. extraction-logic → map argument structure + sentence frames
4. extraction-style → build the style corpus
Academic writing
1. meta-writing → draft sections using My Data + Knowledge + PDF + Web
2. meta-review → multi-reviewer improvement using logic+vocab extractions
3. meta-styling → revise draft against the style corpus
4. meta-rewriting → one-shot style transfer from reference paper
Autonomous research
1. agentic-research → iterative data analysis + literature grounding
2. meta-writing → write up findings with proper citations
3. meta-review → review and improve the manuscript
Output Structure
Knowledge_{topic}/ # Cited knowledge claims
├── index.md
├── Chen2024.md
└── Kim2023.md
Logic_{topic}/ # Structure + logic + frames
├── index.md
└── Weber2021_logic.md
Vocab_{topic}/ # Word inventory + glossary
├── index.md
└── Weber2021_vocab.md
Style_{topic}/ # Style data banks
├── index.md
├── Weber2021_style.md
└── cross_section_matrix.md
Rewrite_{topic}/ # One-shot style transfer output
├── blueprint.md
├── gap_analysis.md
├── rewritten_draft.md
└── session_log.md
Review_{YYYYMMDD_HHMMSS}/ # Draft review reports
├── input.md
├── reviewer_1.md
├── reviewer_2.md
├── synthesis.md
└── improved_draft.md
Research_{topic}/ # Agentic research output
├── world_model.json
├── analysis/
└── report.md
Version History
v1.0.0 (Current)
- Renamed
Meta_researcher → METASCI; plugin name meta-researcher → metasci
- Reorganized skills into four categories: persona, writing (incl. review/proofreading), extraction, research
- Documented persona skills (virgil, beatrice, socrates, picasso, feynman-digest)
- Slimmed feynman-digest; fixed
.claude-plugin/plugin.json skill paths and removed the dead top-level plugin.json
v0.7.0
- Added meta-rewriting skill (one-shot style transfer pipeline)
- Renamed skills for consistency: extraction-{target} + meta-{action} convention
- Updated all cross-references across skills, CLAUDE.md, and README
v0.6.0
- Added agentic-research skill (autonomous data-driven scientific discovery)
- Upgraded meta-writing to v0.2.0: My Data/Knowledge separation, universal (non-domain-specific), writing.local.md config, citation-and-verification, Glob support
v0.5.0
- Added meta-review skill (multi-reviewer draft improvement with four academic writing principles)
v0.4.0
- Added extraction-logic skill (structure, argument logic, sentence frames)
- Added extraction-vocab skill (exhaustive POS extraction, technical glossary)
- Added CLAUDE.md project instructions
v0.3.0
- Added meta-styling skill (Mode A extraction, Mode B revision)
v0.2.1
- Added reference verification to meta-writing
v0.2.0
v0.1.0
- Initial extraction-knowledge skill
License
MIT
Author
KKH
