Source-grounded academic writing for LaTeX + BibTeX. Turns the agent from an eager intern that invents citations into a co-author where every claim traces to a real source.
60 seconds · macOS, Linux, Windows · Node 18+
or, as a Claude Code plugin
Every claim traces to a real source. Never invent a citation, DOI, quote, or quantity. Never overclaim beyond the evidence.
In code, a hallucination breaks the build and you find out. In a manuscript, a hallucinated citation looks correct — and survives to peer review, or print. Same model, same prompt; the kit changes the rules.
The prompt: “Write the introduction.”
Writes fluent prose immediately — which is exactly the danger…
Questions, gathers evidence, drafts in scope, and verifies…
Same model · same prompt · different rules
Every substantive piece of writing walks the same five-step loop. The advisory rules are read at session boot; the deterministic hooks enforce the non-negotiable parts — a dangling \cite or a fake-shaped DOI is blocked, not merely discouraged.
Restate what the section must establish as a claim a reader could dispute. "Write the intro" becomes "Establish that tool-call hallucination in multi-turn agents is understudied and our gate closes that gap."
CLAUDE.md → The Core Loop
Gather the sources that support the claim before drafting. If the evidence is not in the library (references.bib + sources/), say so. Never draft around a citation you intend to "find later."
agent_docs/citation-discipline.md
Write the smallest passage that makes the point, matching the surrounding voice — tense, person, hedging level, one term per concept. Off-thesis tangents are logged under Not Now, not smuggled in.
agent_docs/academic-style.md
Citations resolve → quotes match verbatim → claims are supported → numbers are consistent → cross-references resolve → it compiles. Placeholders are reported, never left silently embedded.
CLAUDE.md → Verification
Every non-trivial claim carries a \cite{key} resolving to a real references.bib entry. A sentence is cited, the author’s own argument stated as such, or common knowledge — nothing else belongs yet.
block-fabrication + citation-gate hooks
“Would a skeptical Reviewer 2 sign off on this sentence?” — the question every section ends with.
Every piece is plain markdown or shell. Read every line in a sitting. The hooks aren’t docs — they’re a contract, and the bench proves it.
Invoke with /skill-name, or let Claude Code load them automatically via semantic matching.
/claim-checkWalk every claim, classify cited / own / unsupported, check the citation licenses the verb.
/citation-auditBibliography health — dangling cites, orphan entries, malformed DOIs, \ref↔\label.
/peer-reviewFull simulated referee report — runs the peer-reviewer + integrity-reviewer agents.
/literature-reviewSynthesize related work from your own library — thematic, gap-driven, real citations only.
/manuscript-cycleOrchestrator — outline → ground → draft → verify → review → revise, halting on any gate.
/response-to-reviewersPoint-by-point letter — quote, change, location; never claims an unmade change.
The analogue of stack templates — discipline-specific reporting standards, nomenclature, and structure that supplement the general docs.
Reproducibility (seeds, compute, decoding), baselines & ablations, significance over seeds, eval contamination.
IRB/IACUC ethics, wet-lab reproducibility (RRIDs, replicates), figure integrity, GEO/SRA deposition.
Preregistration & the replication crisis, construct validity, qualitative coding, effect sizes over bare p.
Prospective trial registration, effect measures (ARR, NNT, ITT), allocation concealment, COI disclosure.
“Don’t invent citations” in a prompt is advisory — the model forgets it three turns later, and a hallucinated DOI looks correct. The kit makes the discipline deterministic: a fake-shaped DOI is blocked by a hook, a dangling \cite fails the gate, and the Literature Vault gives every claim a real source to trace to.
“In code, a hallucination breaks the build and you find out. In a manuscript, a hallucinated citation looks correct — and survives to peer review, or print.”
Read the full rationale
One source. Many tools.
Edit one CLAUDE.md, run ./scripts/convert.sh, and derive a cross-tool AGENTS.md (+ Cursor / Windsurf / Aider).
Sixty seconds, either way. The next manuscript section your agent drafts will be shaped by the cardinal rule.