citation-coverage-audit
Citation Coverage Audit
Audit whether a draft paper cites the works it should cite. This skill is about reference completeness: classic foundations, closest prior work, current concurrent work, benchmark/dataset/method attributions, and reviewer-sensitive omissions.
Use this skill when:
- the user has a draft paper but suspects the reference list is incomplete
- a related work section may miss classic or recent work
- the paper needs to cite concurrent arXiv/OpenReview papers before submission
- the paper makes novelty claims that require stronger citation support
- the user wants to reduce reviewer complaints like "missing important related work"
Do not use this as a replacement for citation-audit. Use citation-audit after this skill to verify that added BibTeX entries and citation keys are correct.
Pair this skill with research-project-memory when missing citations affect novelty, baseline, related-work, or reviewer-risk tracking.
Skill Directory Layout
<installed-skill-dir>/
├── SKILL.md
└── references/
├── coverage-taxonomy.md
├── discovery-protocol.md
├── memory-model.md
└── report-template.md
Progressive Loading
- Always read
references/coverage-taxonomy.mdandreferences/discovery-protocol.md. - Read
references/report-template.mdwhen writing the audit output. - Read
references/memory-model.mdwhen saving or reusing topic-specific citation coverage knowledge. - Use current web search for recent and concurrent work. Do not rely on static memory for "latest" or "recent" citations.
Core Principles
- Missing close citations are submission risks. Treat them as reviewer-facing problems, not bibliography trivia.
- Separate classic foundations, closest prior work, and recent concurrent work. They serve different rhetorical and reviewer functions.
- Search from the paper's claims, not just keywords. A method paper, benchmark paper, theory paper, and empirical study need different citation coverage.
- Prefer primary sources: official papers, proceedings pages, arXiv, OpenReview, ACL Anthology, PMLR, CVF, ACM, IEEE, Springer, Semantic Scholar, DBLP, and authors' project pages.
- Do not pad the bibliography. Recommend citations only when they support a specific sentence, section, baseline, dataset, method component, theorem, or novelty boundary.
- Mark uncertainty. If a candidate looks relevant but cannot be verified, list it as
needs-author-review.
Step 1 - Define Scope
Identify:
- target venue and year
- topic and subtopic
- paper archetype: method, theory, empirical study, benchmark/dataset, systems, analysis, application, or hybrid
- paper stage: outline, full draft, final submission, rebuttal, camera-ready
- current source format: LaTeX, PDF, Markdown, notes
- search horizon for recent work, such as last 6 months, last 12 months, or since the previous conference deadline
- whether the user wants a report only or also BibTeX/prose patch suggestions
Default search horizon:
- for active ML/AI topics: last 12 months
- for fast-moving topics such as foundation models, diffusion, reasoning, agents, alignment, or multimodal learning: last 6-9 months plus current OpenReview submissions if accessible
- for theory or mature topics: include classics and last 24 months
Step 2 - Extract Paper Claims and Existing Coverage
Read the draft and extract:
- title and abstract
- contribution bullets
- introduction novelty claims
- related work categories
- method components
- theorem assumptions or proof techniques
- datasets, benchmarks, metrics, and baselines
- experiment comparison points
- limitation and scope claims
- current citation keys and BibTeX entries
Build a paper citation map:
## Paper Citation Map
- Main claim:
- Core method/theory:
- Closest related work currently cited:
- Classic foundations currently cited:
- Recent/concurrent work currently cited:
- Datasets/benchmarks currently cited:
- Baselines currently cited:
- Novelty claims requiring citation support:
- Sections with thin citation coverage:
If the draft is incomplete, audit the available claims and mark missing context.
Step 3 - Classify Required Citation Categories
Read references/coverage-taxonomy.md.
For the paper, define required buckets:
- foundational classics
- closest prior work
- direct competitors
- concurrent work
- benchmark/dataset/metric sources
- method/tooling components
- theory/proof technique sources
- negative or limitation-related citations
- surveys or taxonomy papers
- venue-specific or community-standard citations
For each bucket, state why it matters for reviewer perception.
Step 4 - Discover Missing Work
Read references/discovery-protocol.md.
Search iteratively:
- use the paper's own title/claim keywords
- search each method component and baseline
- search closest cited papers forward and backward when possible
- search recent arXiv/OpenReview/proceedings for the topic
- search venue-specific accepted/oral/spotlight papers if the target venue matters
- inspect surveys or benchmark leaderboards for canonical citations
For recent/concurrent work, include access dates and search queries. Search results change over time.
Do not recommend a paper only because it shares keywords. Each candidate must map to a citation role.
Step 5 - Evaluate Candidate Relevance
For each candidate missing paper, classify:
must-cite: close prior work, direct competitor, foundational citation, dataset/metric/source attribution, or work required to support/qualify a novelty claimshould-cite: relevant recent work, adjacent strong baseline, useful survey, or important related line reviewers may expectoptional: context or breadth citation that improves related work but is unlikely to affect reviewdo-not-cite: keyword match but not scientifically relevantneeds-author-review: relevance plausible but not enough evidence to decide
For every must-cite or should-cite, specify:
- where to cite it
- what claim it supports or qualifies
- how it changes novelty framing
- whether it needs BibTeX
- whether it suggests new experiments or baselines
Step 6 - Check Novelty and Reviewer Risk
For each novelty claim, ask:
- Is the closest prior work cited?
- Is the difference stated correctly?
- Would a reviewer expect this missing paper?
- Does the missing paper weaken the novelty claim?
- Does it require a new baseline, ablation, theorem comparison, or discussion?
Risk levels:
blocking: missing citation could make a novelty claim look false or unethicalhigh: reviewer likely complains about missing related workmedium: omission weakens positioning but probably not fatallow: useful completeness citation
Step 7 - Produce Coverage Report
Use references/report-template.md.
The report should include:
- paper citation map
- search strategy and search date
- candidate missing citations
- must-cite and should-cite recommendations
- exact insertion points
- novelty-framing updates
- baseline/experiment implications
- unresolved author decisions
If the user wants a saved report and gives no path, use:
docs/reports/citation_coverage_audit_YYYY-MM-DD.md
Step 8 - Optional BibTeX and Text Patches
If the user asks for edits:
- add BibTeX only from verified metadata
- insert citations at the smallest safe location
- do not make broad related-work rewrites unless asked
- do not cite papers you have not verified enough to identify
- after edits, run or recommend
citation-auditto verify keys and metadata
When exact prose is uncertain, write a suggested sentence in the report instead of editing the paper.
Step 9 - Update Citation Coverage Memory
Read references/memory-model.md.
When a topic scan was performed:
- update
.agent/citation-coverage/topics/<topic-slug>.md - update
.agent/citation-coverage/project-coverage.md - record search queries, access dates, candidate papers, and decisions
Memory must separate:
- canonical papers for the topic
- recent/concurrent papers with access dates
- papers already cited
- papers intentionally not cited
- unresolved author decisions
If the project uses research-project-memory, also update:
memory/risk-board.md: missing closest-work, classic, benchmark, baseline, or concurrent-work risksmemory/action-board.md: add-citation, check-bibtex, add-baseline, revise-novelty, or author-review actionsmemory/claim-board.md: novelty claims that must be weakened or qualifiedmemory/evidence-board.md: citation evidence for claims when a paper directly supports or limits a claimpaper/.agent/paper-status.md: related-work sections or insertion points that need edits
Use needs-verification for candidate papers not yet read deeply enough.
Final Sanity Check
Before finalizing:
- classic, closest, and recent/concurrent work were considered separately
- search queries and dates are recorded
- every recommended paper has a citation role
- must-cite recommendations include insertion points
- novelty risks are explicit
- uncertain candidates are not presented as verified
- any edits are followed by a recommendation to run
citation-audit
More from a-green-hand-jack/ml-research-skills
project-init
Initialize an ML research project control root. Use for paper/code/slides repos, shared memory, GitHub Project alignment, agent guidance, worktree policy, and lifecycle handoffs.
37project-sync
Sync verified code-side experiment results into paper memory. Use when logs, reports, run docs, or user-confirmed metrics should become paper-facing evidence.
36add-git-tag
Create annotated Git milestone tags. Use when completing a phase, releasing a version, or marking a research checkpoint.
36update-docs
Refresh project documentation after code changes. Use after implementing features, changing behavior, or preparing a milestone commit.
36init-latex-project
Initialize a LaTeX academic paper project. Use for new conference or journal papers needing templates, macros, venue preambles, and writing guidance.
36new-workspace
Create Git branches or worktrees for research code and paper versions. Use for experiments, baselines, rebuttal fixes, arXiv/camera-ready branches, and worktree memory.
36