Format bibliography

Normalize and reformat a project's .bib bibliography

Bring a bibliography into a single, consistent house style through the web-latex-mcp MCP tools. Four passes, each one the collaborator asked for:

1. Deduplicate — find entries that are the same paper (often an arXiv preprint and its published version) and merge them down to one.
2. Consistent cite keys — rename to firstauthorYEARtag (e.g. smith2024fastnet), and update every \cite{…} in the draft to match.
3. Harmonize venues — make every proceedings/journal name use one style: short (CVPR) or long (Computer Vision and Pattern Recognition), the user's choice — not a mix.
4. One field policy — either strip or systematically add the optional fields (url, doi, pages, …), consistently across every entry.

Unlike verify-citations (read-only), this skill rewrites the .bib and the .tex. Treat it accordingly.

This skill writes — permission first, compile as the guardrail

• Get explicit approval before any write. Agree the policy with the user, preview the changes (renames, merges, field edits), and only then apply. Never reformat a bibliography unprompted.
• .bib is protected. Every .bib write goes through edit_file/write_file with confirmBibEdit: true — set it only after the user has approved the change.
• The compile is the safety net. The project must compile cleanly before you start and after you finish, with no Citation … undefined warnings — that warning is the tell-tale of a key rename that didn't reach the .tex. Reformatting deliberately changes how the bibliography renders (e.g. CVPR vs the full name); what must not change is the set of papers cited or the project's ability to compile.
• Prefer DBLP as the source of truth. When a value needs to be added or corrected (a venue's canonical name, a missing doi/pages, an author list), pull it from DBLP via search_references rather than inventing it — same principle as the rest of this server: bibliographic data comes from DBLP, not the model. Mechanical transforms (renaming a key, deleting a field, fixing whitespace) you may do directly.
• Idempotent. Skip entries already carrying a % formatted-by-claude: comment (see the marker section) unless the user asks to redo them.
• Commit/push only when asked — per CLAUDE.md, stop at the diff.

Workflow

Run in order. Stop and report if a step fails.

1. Pick the project. If the user didn't name one, list_projects and ask which.
2. Sync. project_sync so you reformat the current sources.
3. Baseline compile. compile and record success + the warning list. If it does not compile cleanly to begin with, stop and tell the user — you won't be able to tell your changes apart from a pre-existing break, and you can't trust the "no undefined citations" check.
4. Read everything. list_files for bib and tex; read_file each. Parse the .bib into entries (type, cite key, all fields) and scan the .tex for every citation command and key (see Renaming keys for the command list). Skip entries already marked % formatted-by-claude:.
5. Agree the policy with the user. Use AskUserQuestion to settle, in one go:
   • Venue style — short acronyms or full names?
   • Optional fields — for each of url, doi, pages (and isbn/publisher/address/eprint): strip them, add-where-missing (from DBLP), or leave as-is?
   • Cite-key scheme — confirm firstauthorYEARtag (the default below) or take their variant.
   • Duplicates — confirm the merge strategy (which version wins; see Detecting duplicates).
6. Compute and preview. Produce the full change set without writing yet: the key-rename map (old → new), the merges (which keys collapse into which), and the per-entry field changes. Show it to the user and get a clear go-ahead. For a big bibliography, summarize and show a representative sample.
7. Apply. With approval: edit the .bib entries (confirmBibEdit: true), and propagate every key rename/merge into the .tex in the same pass (see Renaming keys). Re-fetch added field values from DBLP. Do all edits within the one project so the per-project mutex serializes them.
8. Recompile and verify. compile again. It must succeed with zero Citation … undefined warnings and the same count of distinct papers cited as the baseline. If anything broke, fix it; if you can't, discard and report.
9. Mark. Add a % formatted-by-claude: comment to each entry you reformatted (see the marker section).
10. Review. Show the diff (both .bib and .tex). Do not commit/push unless asked.

Citation key scheme: firstauthorYEARtag

Lowercase ASCII, letters and digits only. Three parts, concatenated with no separators:

• firstauthor — the family name of the first author, lowercased, diacritics stripped (é→e, ü→u), spaces/hyphens/apostrophes removed. For a multi-token surname keep the principal name token (e.g. van den Oord → oord, Le Cun → lecun).
• YEAR — the 4-digit publication year.
• tag — the paper's acronym / method name if it has one, else the first meaningful word of the title:
  • Many titles are Acronym: full title — take the part before the colon (FastNet: … → fastnet).
  • Otherwise look for a distinctive mixed-case or all-caps token in the title (ResNet, BERT).
  • No acronym → the first word of the title, skipping a leading article (A/An/The). Lowercase, letters/digits only.

Example: Smith et al., 2024, "FastNet: A Fast Approach…" → smith2024fastnet.

Collisions: if two distinct papers reduce to the same key, append a, b, c, … to disambiguate (smith2020deep, smith2020deepb). Two entries that collide and are the same paper are duplicates — merge them instead (next section).

Detecting duplicates

Two entries are the same paper when any of these hold (normalize first — strip {}, LaTeX, case, whitespace):

• identical doi, or
• identical normalized title and same first-author surname (year may differ by the preprint/published gap), or
• one is an arXiv @misc/preprint and the other its published @inproceedings/@article of the same title.

To merge: keep the published version over the preprint (confirm with the user when unsure), repoint every \cite of the dropped key to the kept key, and delete the dropped entry. Note the merge in your report. When both carry useful fields, prefer the kept entry's values but fill gaps from the other (or from DBLP).

Harmonizing venues

Apply the user's chosen style (short or long) to every booktitle/journal. Treat these as the same venue in either direction — DBLP's venue field is the canonical short form; derive the long form from it:

• CVPR = (Proceedings of the IEEE/CVF Conference on) Computer Vision and Pattern Recognition
• ICCV = International Conference on Computer Vision · ECCV = European Conference on Computer Vision
• NeurIPS / NIPS = Advances in Neural Information Processing Systems (renamed NeurIPS in 2018)
• ICML = International Conference on Machine Learning · ICLR = International Conference on Learning Representations
• AAAI, IJCAI, ACL, EMNLP, NAACL, KDD, SIGIR, WWW, SIGGRAPH, MICCAI; journals TPAMI, IJCV, JMLR

The list isn't exhaustive — for anything not on it, confirm the canonical name against DBLP rather than guessing. Don't change which venue an entry claims; only its spelling/style. (If an entry's venue looks wrong, that's a verify-citations job, not this one — flag it, don't silently "fix" it here.)

Field policy (url, doi, pages, …)

Apply one rule per field across the whole bibliography, per the user's choice in step 5:

• Strip — remove the field from every entry (mechanical; common for url/note/eprint when a venue forbids them).
• Add where missing — fill the field from the DBLP record for that paper (so doi/pages/url come from the API, not the model). If DBLP doesn't have it, leave it absent and note which entries couldn't be completed.
• Leave as-is — don't touch it.

Keep this surgical: change only the fields the policy covers, plus the cite key. Don't rewrite an entry wholesale unless the user asked you to re-fetch it from DBLP.

Renaming keys without breaking \cite

This is the dangerous part: a renamed key that isn't updated everywhere becomes an undefined citation.

• Build the complete old → new map (including merges, where several olds map to one new).
• Replace keys in every citation command across all .tex files: \cite, \citep, \citet, \citeauthor, \citeyear, \autocite, \parencite, \textcite, \footcite, \Cite…, \nocite, and friends — including comma-separated lists (\citep{a,b,c}: rename each key independently) and ignoring the optional [...] arguments. Also update any crossref = {oldkey} inside the .bib.
• Rename one key at a time, both sides together (its .bib entry and all its .tex uses), so the cite graph is never half-renamed. Watch for accidental substring hits — match whole keys, not fragments.
• The post-compile check is the proof: zero Citation … undefined warnings, and no orphaned entries. If the user only wants the .bib touched and not the .tex, then do not rename keys — tell them renames require editing the draft.

The % formatted-by-claude marker

After reformatting an entry, add one comment line immediately above it so a later run skips it:

% formatted-by-claude: key smith2024fastnet on 2026-06-29
@inproceedings{smith2024fastnet,
  ...
}

• A % line in a .bib is a comment — BibTeX/biber ignore it, so the marker never affects the output.
• Record the final cite key and today's date.
• It's a .bib write like any other: edit_file with confirmBibEdit: true.
• This marker is independent of verify-citations' % verified-by-claude: line; an entry can carry both, and renaming its cite key doesn't invalidate the verified marker (that one records the DBLP key).
• Step 4 skips entries that already have this marker — that's what makes re-runs cheap and stops a second invocation from re-formatting the same entries.

After you finish

Report concisely: how many entries were deduplicated (and into which keys), how many keys were renamed (with the corresponding .tex updates), the venue style and field policy applied, and any entries DBLP couldn't complete. Confirm it still compiles with no undefined citations, and point the user at the diff for both the .bib and the .tex. Remind them nothing is committed or pushed until they ask.