Claude Language Assessment Research Kit (CLARK)

Research skills for the language-assessment and applied-linguistics community. From the literature watch to the pre-submission check, each skill packages a complete, integrity-first workflow — integrity-first meaning nothing is ever invented, nothing is silently changed, and the AI-use declaration your venue expects is a built-in reminder.

CLARK — the whole research workflow: operated by Claude, checked at every step, decided by you. It doesn't read your papers for you and won't write your argument. It runs the machinery around the scholarship: renders your citations perfectly, audits every reference against live records (dead DOIs, mis-attached works, retractions, superseded preprints), watches the field for new papers, recomputes the statistics in your manuscript, and reads your draft through your target journal's own reviewer criteria — with a second, different-vendor AI on call to re-judge any of it. What stays yours is everything that makes it research: the reading, the relevance calls, the interpretation, and the claims you put your name to. Claude is the operator and the checker — never the author.

Every skill also works on its own — run just the citation check, or just a review, in any order. And a skill is a plain instruction file, not an app: the full machinery lives in Claude Code (renders, live Zotero, scheduled radars), and a skill can travel to lighter setups — upload one to a chat interface and Claude follows the same workflow within that interface's limits.

CLARK's home field is language assessment — the reviewer criteria, venue guides, and examples are built from that field's own materials, not from a generic research template. But nothing in the machinery is assessment-specific: if you work anywhere in applied linguistics (SLA, CALL, EAP/L2 writing, corpus research…), the pipeline works as-is — your journals join the reviewer-guides file on demand. The rest of this page documents the citation skills — the toolkit's foundation.

🧭 New: CLARK Field Guides — expert-curated domain packs the skills consume: a domain's venues, reviewer criteria, and review vocabularies, curated by a named expert and citable with its own DOI (the skills are the verbs; the guides are the nouns). The founding guide (Automated Writing Evaluation) is in preparation; experts from any corner of the field are invited to curate one — listening assessment, item writing, young learners… — at whichever of the three contribution rungs fits, from "cite my published work" to full co-curatorship. Spec, provenance rules, and registry: field-guides/.

The citation pipeline

You write [@cheng2026] in a plain-text draft. You get a Word document with "(Cheng et al., 2026)" formatted perfectly, plus a matched, alphabetised reference list — and a typo'd citation can't hide. No more hand-formatting citations, no more reference lists that drift out of sync with the text.

This repo packages a tested workflow built from free tools plus an AI operator that runs it for you, with a plain-language guide written for people who have never used any of them:

Claude Code runs the fiddly parts — terminals, installs, render commands, warning-reading — so you never have to figure out Pandoc or a command line. The few things it can't do for you (a handful of clicks inside the Zotero app) it walks you through one checkpoint at a time, with screenshots.

Get started (~30 min, mostly automated)

1. Click the green Code ▾ button above → Download ZIP → unzip. (No git needed.)
2. Install Claude Code (desktop app or CLI).
3. Open the unzipped folder in Claude Code and say: "Read skills/zotero-citations/SKILL.md and set up automatic citations for me."

Claude then does the computer work itself — checks/installs Pandoc, places the files, renders the demo so you see it working first — and guides you click-by-click through the Zotero steps, confirming each one worked before the next. From then on, daily use is three moves: add a paper to Zotero → cite by [@key] → say "render my draft." And that's the standing habit for everything in this kit: at any step — Zotero setup, linking Obsidian, the audits — ask Claude for a step-by-step guide, or simply ask it to set it up for you.

The operator's manual: docs/01 (setup, with screenshots), docs/02 (daily use), docs/03 (every gotcha we hit, with fixes), and docs/04 — the citation-integrity rules: how this workflow guarantees no fabricated, no "Frankenstein", and no orphaned references. These are the steps and rules Claude itself follows — read them to see exactly what's happening under the hood, or to do any step yourself. The discipline is author-agnostic: however the prose was drafted, references must resolve, numbers must trace to their tables, and described methods must match the instruments actually included.

Level 2 (recommended once the basics work): give Claude a live connection to your Zotero library — it verifies records against Zotero itself, finds papers by meaning (semantic search), and spots duplicates/missing DOIs while you write. Under the hood this is an MCP connection (Model Context Protocol — the open standard for giving AI assistants live access to tools; we consume the open-source zotero-mcp server). Claude installs this for you: say "Read docs/05 and connect yourself to my Zotero library" — your only manual part is one checkbox in Zotero. Details & what it adds: docs/05. Reversible; everything stays on your computer.

Writing in Obsidian (or want to)? Your drafts are plain Markdown, so Obsidian plugs in with zero pipeline changes — docs/06 is the two-click setup (open your project folder as a vault), the optional citekey-autocomplete plugins, and a render-tested table of what happens if Obsidian-only syntax (like %%hidden comments%% — which Word prints!) leaks into a manuscript.

Running a literature review? docs/07 is the living-review system: a matrix you approve row by row, generated Obsidian views (per-paper notes, theme/method hubs, a dashboard), and an optional cross-vendor AI audit of your own judgements — a second rater, not a ghostwriter; flag-never-guess throughout.

About to submit a paper? docs/08 is the pre-submission review of your own manuscript: page-anchored findings and recomputed statistics, a triage list you adjudicate before anything is written up, a prioritized revision plan with a 0–100 estimate, and a compliance check against the journal's current author instructions.

How it all fits together: see docs/flowchart.md — the diagrams render right here on GitHub. A designed, printable version is docs/flowchart.html (GitHub shows HTML as code — after downloading the ZIP, just double-click that file to open it in your browser).

Changing the citation style (APA 7th is just the default)

The output format is controlled by one file: starter-kit/style.csl. We ship the official APA Style 7th edition. To use a different style — another edition, Chicago, MLA, or a specific journal's house style:

1. Search the free Zotero Style Repository for your style or journal name.
2. Download the .csl file and save it as style.csl, replacing the one in starter-kit/.
3. Re-render. Same draft, new format — the in-text citations and the reference list both follow the new style automatically.

Your draft never changes: [@cheng2026] stays [@cheng2026]; only the style file decides how it looks. (Tip: if you work with two journals, keep two folders with different style.csl files.)

What Claude Code does here (the built-in operator)

skills/zotero-citations/SKILL.md teaches Claude this workflow's five jobs: guided onboarding (the Get-started setup above), per-project setup (copies the starter-kit into any new project and verifies the sample render), render-and-explain (runs the render, translates every warning into plain language), library health-check (missing years/DOIs, duplicates, stale exports), and a pre-submission integrity pass (every citation traced to a real record, text ↔ reference list reconciled both ways) — always flagging anything it can't verify rather than guessing, under the full rule set in docs/04-citation-integrity.md. The judgement calls — what to read, what to cite, what a source says — stay yours.

The integrity pass is also a standalone skill — skills/citation-integrity/ — callable on its own ("check my citations") even without the render pipeline. It runs three library checks in starter-kit/: check_retractions.py (Crossref / Retraction Watch), resolve_check.py (does every DOI resolve to a real record, and does that record match your entry — catching dead DOIs and mis-attached / "Frankenstein" references), and vor_check.py (which of your cited preprints now have a published version of record). All are standard-library Python, no install.

To use the skills in another project of yours: copy this whole repo folder into that project (keeping it together — the skills reference the shared docs/ and starter-kit/ folders by relative path, so separating a skill folder from the repo breaks it), then invoke by path: "Read <folder>/skills/zotero-citations/SKILL.md and set up automatic citations for me." If you place it under the project's .claude/skills/, same rule: the repo folder stays intact. (.claude is a hidden folder; create it if it doesn't exist — Mac Finder shows hidden files with Cmd+Shift+.; Windows Explorer: View → Show → Hidden items.) Full setup patterns and the session habits that keep public tools and private research apart: docs/00.

Requirements & versions

• Zotero (free; Mac/Windows/Linux — no account needed for this workflow) + Better BibTeX plugin (free)
• Pandoc (free; Mac/Windows/Linux)
• Claude Code — the operator (subscription or API). (The underlying pipeline is all free software; the manual in docs/ covers every step if you ever work without Claude.)
• Any plain-text editor for your draft (Markdown) — including Obsidian as a comfortable writing home (docs/06)

Unsure whether your setup is complete? Ask Claude to "run the CLARK doctor" — one command (starter-kit/clark_doctor.py) checks everything in this list, your project's files, and whether your audit keys are in place, without changing anything.

Last verified: July 2026 against Zotero 9, current Better BibTeX, Pandoc 3.10. These tools are open-source and update regularly — a menu may move, but the concepts don't change; each guide step links the official docs as the always-current source.

License

MIT for everything in this repo, except starter-kit/style.csl (the APA style file), which comes from the CSL project and keeps its own Creative Commons BY-SA license (noted inside the file). See LICENSE.

A note on the name: CLARK was developed under the working name "CLAS" before its public release, so you may still see that name in the earliest commits, the v1.0.0 changelog entry, and one internal generation marker (clas-literature-review) kept deliberately for backward compatibility with vaults created before the rename.