tannhatcms/jarvis/ams-citation-manager

Citation Manager

Citations are the paper's supply chain: every claim traceable to its source. Managing them means collecting sources systematically, organizing them so you can find anything, formatting consistently, and — critically — verifying that every citation is real and says what you claim.

Overview

The workflow: capture sources as you read (never "I'll find it later"), store them with notes on why they matter, cite-while-you-write with keys instead of hand-typed references, and generate the bibliography from the library. The integrity rule: every in-text citation matches a real source you've read, and every bibliography entry is cited. Reference managers automate formatting; they don't automate honesty.

When to use

• Writing papers, theses, reports, or any document with references.
• Building a personal research library over time.
• Collaborating: shared libraries keep everyone's citations consistent.
• Cleaning up a manuscript's references before submission.

Core concepts

• Capture at reading time: save the full citation + PDF + your notes the moment you read

something useful. "I'll relocate it later" fails reliably.

• Cite keys: stable keys (smith2024attention) used in drafts; the formatted reference is

generated, never hand-typed. Renumbering-proof.

• Annotated library: each entry with a one-line note — what it claims, why you cited it.

Future you will thank present you.

• Style consistency: one citation style per document (APA, IEEE, Chicago…), applied by the

tool, not by hand. Switching styles should be one setting.

• Reference integrity: every citation verified — the source exists, you read the relevant

part, it supports your claim. AI-generated citations must be checked individually.

• Deduplication: the same source saved three ways creates triple entries. Merge duplicates;

keep one canonical record.

Practical workflow

1. Set up the library before writing: folder structure or tags by project/topic.
2. For each source read: save citation metadata, the file, and a 2-line note (claim + relevance).
3. Write with cite keys; never type a formatted reference by hand.
4. Before submission: run the integrity check — every key resolves, every entry is cited, no

duplicates, style consistent.

1. Verify AI-suggested citations individually: confirm the source exists and supports the claim.
2. Archive the library with the manuscript version so the reference set is reproducible.

Source note template:
KEY:   smith2024attention
CLAIM: <what the paper shows, in your words>
USE:   <why you're citing it — which argument it supports>
READ:  <which sections you actually read>
LIMIT: <caveats relevant to your use>

Common pitfalls

• Phantom citations: references that don't exist — the classic AI failure mode. Verify every

single one.

• Citation without reading: citing the abstract or a citation-of-a-citation. Read the part you

rely on.

• Hand-formatted references: typed bibliographies that break on every edit. Generate from keys.
• No notes: a library of 500 PDFs with no memory of why each mattered. Annotate at capture time.
• Style drift: mixed formats in one document. Set the style once; let the tool enforce it.
• Uncited bibliography entries: sources in the list that nothing cites (or claims nothing

supports). Audit both directions.