WILLOSCAR/research-units-pipeline-skills

Agent Survey Corpus downloads open-access arXiv survey PDFs about agentic systems and extracts text for local reference. Use it to study how real surveys organize sections, subsections, and evidence-backed comparisons.

Anchor Sheet extracts subsection-specific evidence hooks from draft packs, focusing on numbers, benchmarks, metrics, and limitations. Use it when you want sections to be grounded in cited facts instead of generic summaries.

Appendix Table Writer curates reader-facing survey tables for paper appendices using only in-scope evidence and existing citation keys. Use it when you have evidence packs, anchor sheets, and citations and need publishable tables instead of internal logs.

ArXiv Search retrieves paper metadata from arXiv keyword queries and writes normalized results to JSONL. Use it to build an initial paper set for ranking, taxonomy, or citation workflows, or to import and enrich offline exports.

Argument Selfloop builds a final machine-readable snapshot from current H3 sections, refreshes section hashes, and checks consistency before merge. Use it after section edits when you need an updated argument ledger and manifest.

Artifact Contract Auditor checks whether completed Units and the locked Pipeline have all required output files present, then writes output/CONTRACT_REPORT.md. Use it for mid-run coverage snapshots or final delivery completeness, not provenance checks.

Beamer Compile QA compiles `latex/slides/main.tex` into `latex/slides/main.pdf` and writes `output/SLIDES_BUILD_REPORT.md`. Use it when a Beamer tutorial deck needs a build check and a readable success or failure report.

Beamer Scaffold generates a compile-ready Beamer deck from a finished tutorial and approved module plan. Use it when you need to turn output/TUTORIAL.md into latex/slides/main.tex.

Bias Assessor enriches `evidence-review` extraction tables with lightweight risk-of-bias fields before synthesis. Use it when `papers/extraction_table.csv` exists and you need consistent low/unclear/high ratings plus brief notes.

Chapter Briefs builds internal H2 chapter cards with throughlines, contrasts, and lead-paragraph plans for chapters that contain H3 subsections. Use it when you want a survey outline to read like a paper without adding more headings.

Chapter Lead Writer writes body-only H2 lead blocks that preview a chapter’s comparison lens and connect multiple H3 subsections as one argument. Use it when a section needs a lead paragraph without adding new facts or moving beyond the chapter’s citation scope.

Chapter Skeleton builds a retrieval-informed chapter outline from taxonomy and scope before detailed subsection decomposition. Use it when chapter-level structure should be stabilized first.

Checkpoint Brief summarizes stage artifacts into a human-review block in DECISIONS.md before a workflow checkpoint. It is used to record scope, coverage, and unresolved choices without approving or changing artifacts.

Citation Anchoring checks that citation keys stay in the same ### subsection after polishing. Use it to catch drift that breaks claim-to-evidence alignment.

Citation Diversifier generates an in-scope citation budget report for each H3 so a draft can meet global unique-citation targets without adding new facts. It is used before citation injection when a survey or paper is under-cited or reuses the same sources too often.

Citation Injector applies a citation budget report to an existing draft by adding only in-scope citations and no new facts. Use it when a draft needs more unique citations or less citation reuse to pass a global citation gate.

Citation Verifier generates BibTeX from paper notes and writes matching verification records for each reference. Use it when you need traceable citations for prose or LaTeX, including offline-generated records for later verification.

Claim Matrix Rewriter rewrites outline/claim_evidence_matrix.md as a bullet-only index driven by evidence packs instead of outline placeholders. Use it when you need the matrix to preserve the legacy claim/evidence format while reflecting evidence-first inputs.

Claim-Evidence Matrix builds a section-by-section claim and evidence outline from paper notes and a survey outline. Use it before prose writing to make claims explicit and traceable to sources.

Claims Extractor turns manuscript text into a traceable claim ledger for review work. It records each claim with a source pointer and separates empirical from conceptual claims.

Concept Graph builds a deterministic prerequisite DAG from an approved tutorial spec and writes it to outline/concept_graph.yml. Use it when tutorial concepts need stable ordering before module planning.

Dedupe Rank turns a broad paper candidate pool into a deduped JSONL and a stable core set. Use it when retrieval output needs deterministic deduplication, ranking, and repeatable paper IDs.

Deliverable Self-Loop scores an existing reader-facing deliverable and writes a deterministic PASS/FAIL report with a bounded repair plan. Use it after synthesis to verify section and artifact contracts without rewriting the content.

Draft Polisher edits `output/DRAFT.md` to remove boilerplate, improve coherence, and keep citations anchored to the evidence. Use it when a first-pass draft reads like scaffolding or needs a final polishing pass before review.

Reference-First Template provides a reusable shape for skills built around a lean router, reference files, machine-readable assets, and a deterministic script. Use it when creating or refactoring a skill into this repository's reference-first structure.