Skip to content

README.md

Latest commit

 

History

History
108 lines (68 loc) · 5.17 KB

README.md

File metadata and controls

108 lines (68 loc) · 5.17 KB
Raw

/bibcheck — Many-Agent Bibliography Audit

A verification skill for .bib files. Spawns many narrow-focus agents — one per citation, or one per field — to audit each bibliography entry against canonical sources (DOI, journal page, author working paper). Catches the silent errors that pass spell-check but kill peer review: mixed-up entries, wrong years, fabricated DOIs, journal misattributions.

Why this exists

A single agent asked to audit 80 citations in one pass tends to drift. Early entries get careful treatment; later entries get pattern-matched. Whatever you call it — gradient decay, attention dilution, fatigue — the practical effect is that late-batch citations are checked less rigorously than early-batch citations, and the user cannot see which is which.

/bibcheck flips the structure. Each agent gets one small task, full attention budget, and a parallel sibling for the next entry. The bottleneck moves to orchestration, which is what cheap parallel agents are for.

This is the same principle behind /referee2 (fresh agent reads cold), /split-pdf (parallel chunks instead of one giant PDF), and the package-audit experiment (96 zero-discretion agents instead of one general-purpose one).

What it checks

For each entry, the audit verifies:

  1. Existence — does the cited paper exist? Locate a DOI or canonical landing page.
  2. Field consistency — do title, authors, year, journal, volume, issue, pages, and publisher all belong to the same paper? (Mixed-up entries are the most common silent error in inherited .bib files.)
  3. Field correctness — wrong year, wrong abbreviation, wrong volume, wrong page range.
  4. Format integrity — parseable BibTeX, no quietly-broken unicode in author names.

What it does not check:

  • Whether the citation supports the claim it's attached to (that's a /referee2 literature audit).
  • Whether the cited paper's findings are correctly characterized in the manuscript text (same).
  • Whether the manuscript is missing citations it should have.

/bibcheck is bounded: are the entries in this .bib accurate descriptions of papers that exist?

Two modes

Per-citation (default)

/bibcheck path/to/refs.bib

One agent per @article{}/@book{}/@incollection{} entry. Each agent fully audits its one entry: identify the paper from canonical sources, cross-check that all fields belong to the same paper, emit a one-sentence description and a corrected entry if needed. A final reviewer agent consolidates and adjudicates.

Best for: catching mixed-up entries (title of paper A paired with authors of paper B).

Per-field

/bibcheck --by-field path/to/refs.bib

One specialist per field — title, year, journal, authors, volume/issue, pages, DOI. Each specialist reads the whole bibliography but only checks its one field. Specialists are launched as separate claude --dangerously-skip-permissions -p subprocesses so they do not share context, then the consolidator joins their outputs.

Best for: catching systematic transcription errors — a journal name consistently rendered wrong, working-paper years leaking into published-paper entries, swapped volume/issue numbers.

The two modes are complements. If a manuscript is going somewhere it cannot afford a citation error, run both.

Outputs

<bib_dir>/bibcheck_<timestamp>/
  ├── input.bib              # copy of source
  ├── entries/               # split entries
  ├── reports/               # per-agent JSON
  ├── bibcheck_report.md     # consolidated audit
  └── corrected.bib          # drop-in replacement

The skill never auto-overwrites your source .bib. You review bibcheck_report.md, then move corrected.bib into place yourself.

Arguments

Argument Effect
<file> Path to a .bib (or a .tex whose \bibliography{} points to one)
--by-citation Per-citation mode (default if omitted)
--by-field Per-field mode
--max-parallel N Cap concurrent subagents/subprocesses. Default 8

Cost and rate-limit notes

/bibcheck issues one WebSearch (often 2–3) per citation in per-citation mode, plus an aggregator pass. For an 80-entry bibliography, expect ~80 light agent runs plus one consolidation. Per-field mode is heavier — each specialist does N web searches across the whole bibliography. Default --max-parallel 8 keeps things sane on a Max plan; bump it if you have headroom or lower it if you are getting throttled.

When to use which mode

Situation Mode
Inherited .bib from a coauthor or RA per-citation first
Pre-submission audit on your own manuscript per-citation, then per-field on the entries flagged unverifiable
Worried about systematic transcription errors (e.g., one journal abbreviated wrong everywhere) per-field
3+ entries came back unverifiable re-run those specifically with per-field

Related skills

  • /referee2 — broader audit of an artifact, including a literature-review check (citation-claim fidelity)
  • /split-pdf — if your starting point is a PDF rather than a .bib, run this first to extract the references section