You have a research idea. Now you need a study design,
an analysis plan, a critical review,
and someone to tell you your methodology has holes. Scholark is that someone, except it doesn't
have feelings, and it works at 3 AM. It won't replace your taste or judgment. Nothing can. But it will
make sure you've thought of everything before you bet months of work on an idea.
SYSTEM MANIFEST
WHAT
A Claude Code plugin for HCI researchers. 8 skills and 3 specialized agents that take you from vague idea to pre-submission-ready manuscript. No hallucinated methodology, no hallucinated readability scores, no hallucinated power analyses, no hallucinated citations.
HOW
Two research agents work in parallel: one generates ideas with creative abandon, the other tears them apart with adversarial rigor. The surviving ideas get formalized into proper study designs. A third, read-only agent fact-checks every citation against the actual source paper before you submit.
SETUP
Clone, point Claude Code at the directory, done. No build step, no compilation. Pure markdown and JSON, plus small optional Python helpers (readability and power calculations) that offer to install themselves on first use.
COST
Free. MIT licensed. The plugin is declarative markdown. There's nothing to monetize even if we wanted to.
FOR WHO
HCI researchers, interaction designers, UX scientists, anyone who designs studies with humans. Built by a professor who reviews papers for a living.
// SKILLS
/scholark:research-brainstorm
Dispatches both research agents in parallel. Returns ranked ideas with risk assessment and contribution-type mapping.
/scholark:study-design
Conversational study design formalization. IVs, DVs, design type, sampling strategy, full procedure.
/scholark:analysis-plan
Pre-registration-ready statistical analysis specification. Real power analysis via a bundled calculator, assumptions, exclusion and stopping rules, and an optional OSF-style pre-registration draft.
/scholark:study-validator
Completeness checklist with severity-rated gaps. Catches what you missed before reviewers do.
/scholark:literature-blind-spots
Accepts .tex drafts (.md as fallback), searches for citation gaps across 4 academic databases, outputs an HTML reading plan with verified DOIs.
/scholark:paper-review
Pre-submission review against common rejection patterns at top HCI venues. Catches structural and argumentative issues reviewers actually flag at CHI, CSCW, UIST, DIS, TOCHI, IJHCS.
/scholark:prose-tighten
Tightens .tex prose without losing field-specific terminology. Splits long sentences, cuts fluff, fixes nominalizations. Asks before touching technical terms. Reports a real Flesch diagnostic via textstat. Never hallucinated.
/scholark:citation-integrity
Audits every \cite{} in your .tex (or .md) manuscript. Spawns one read-only verifier agent per cited paper, each reading the source in full from a local context folder. Surfaces Wrong, Misleading, Partial, Unverifiable, and Not audited citations with suggested fixes. Offline by design. No web, no databases, just the papers you provide.
// AUTONOMOUS AGENTS
research-ideator
RUNS ON SONNET // WEB + ACADEMIC SEARCH
The optimist. Generates research ideas by combining web trends with academic literature. Creative, broad, ambitious. Will suggest things that make you say "wait, could that actually work?"
research-critic
RUNS ON SONNET // ACADEMIC SEARCH ONLY
The pessimist. Reviews ideas using only academic sources. Web search is deliberately disabled, so it can only cite what's published. Finds the holes, flags the risks, saves you from Reviewer 2.
citation-verifier
RUNS ON SONNET // READ-ONLY // NO WEB, NO MCP
The fact-checker. Reads one cited paper from your local context folder and judges whether the manuscript's wording matches what the source actually says. Dispatched in parallel by /scholark:citation-integrity, one agent per cited paper. Tool-restricted to Read so it cannot wander or fabricate. Returns Verified, Partial, Misleading, Wrong, or Unverifiable verdicts with suggested fixes.
SCHOLARK-1 // MCP SERVER
Scholark works standalone, but it gets sharper with real papers behind it. Scholark-1 is a companion MCP server
that searches 4 academic databases in parallel (Semantic Scholar, OpenAlex, Crossref, Europe PMC) plus Unpaywall
for open-access PDFs. Real papers, real DOIs, real BibTeX. The brainstorm and literature-blind-spots skills use
it automatically when available.
Step 2: read the HTML report, decide which papers to add
Step 3: /scholark:paper-review (check the strengthened argument holds together)
> PRE-SUBMISSION CITATION AUDIT
Step 1: /scholark:citation-integrity paper.tex (extracts every \cite{}, gates on coverage)
Step 2: convert any missing source PDFs to markdown via the bundled marker_single wrapper
Step 3: one read-only verifier per cited paper, in parallel, returns a severity-sorted audit report
Step 4: apply suggested fixes interactively, per claim, with confirmation
Final step in every flow: You. The human. Reading, thinking, deciding what stays. AI generates options. Taste and judgment are yours.
// FEATURED // ONE PROMPT, A REAL PIPELINE
The most-loved Scholark move. Hand Claude your draft and a single instruction. What looks like one prompt is actually a chain of skills, agents, and database queries working together.
"Read my paper.tex, identify 3 blind spots in the literature, use scholark-1 to find papers that address them, and generate a studyplan.html with clickable DOI links and notes on how each paper relates to what's missing"
> WHAT ACTUALLY RUNS
1. Claude reads the .tex source. 2. The literature-blind-spots skill maps the argument and groups missing citations into 6 categories: foundational work, adjacent fields, methodological precedents, competing frameworks, recent developments, and replication or meta-analysis coverage. 3. For each category, scholark-1 fires three kinds of queries against 4 academic databases (Semantic Scholar, OpenAlex, Crossref, Europe PMC) in parallel: โธ broad keyword search โธ topic-focused search with year filtering โธ targeted metadata lookup for each candidate paper 4. Results are deduplicated by DOI across all sources. 5. When an API throttles, the skill falls back to the open web (Google Scholar, ACM Digital Library, IEEE Xplore) and pulls paper details from publisher landing pages. 6. Every author and title is re-verified against the source database before it lands in the output. Initials are never expanded into full names. Conflicting metadata is flagged for manual check, not silently reconciled. 7. The output is a self-contained HTML reading plan with clickable DOIs and a note on how each paper relates to a specific section of your draft.
One prompt in. Dozens of real tool calls and database queries executed. A reproducible reading plan out, with every citation grounded in published metadata. The LLM provides judgment about what to look for. The tools provide ground truth about what exists. This is the difference between Scholark and a chatbot: there is a real pipeline behind the prompt.
Free and open source. MIT licensed.
Your AI research assistant that actually knows how studies work.