Literature Survey

Literature Survey

When to load this skill

Load this skill when the user is:

• starting a new research paper that needs a literature review
• adding new references to an existing paper
• migrating an unverified BibTeX file to a verified one
• doing a deep-read pass on a set of references and producing per-paper notes
• needing a reproducible record of which references have been verified vs which are still placeholders

Overview

A research paper's bibliography drives credibility. Stale, mis-attributed, or unverified entries embarrass the author at submission time. This skill defines a 5-step workflow for building a verified literature corpus and a deliverables structure that future-you (or a collaborator) can audit and extend.

The five steps:

1. Add a verified BibTeX entry to references/bibliography.bib.
2. Drop the published-manuscript PDF into references/pdf/<citekey>.pdf.
3. Generate a text extraction for AI consumption.
4. Write a survey note notes/survey_<citekey>.md.
5. Append a row to references/_collection_log.md.

Each step is described in detail below, plus the directory layout, the survey-note template, and the collection-log template.

Directory layout

<paper-repo>/
  references/
    bibliography.bib          -- BibTeX (tracked); every entry verified
    _collection_log.md        -- per-paper verification status + corrections + notes
    pdf/                      -- published-manuscript PDFs (gitignored)
      <citekey>.pdf
      <citekey>-supp.pdf      -- supplementary material where applicable
      .txt/                   -- pdftotext -layout extractions (gitignored)
        <citekey>.txt
  notes/
    survey_<citekey>.md       -- per-paper survey note (~30-50 lines)
    README.md                 -- index of survey notes (by section, by affinity)

The references/pdf/ directory and its .txt/ cache should be gitignored; the bibtex + collection log + survey notes are tracked.

Step 1: Verified BibTeX entry

Add the entry to references/bibliography.bib. Required fields: must include venue + page numbers + DOI or arXiv ID. Reject placeholder entries (e.g. "Author, A. (year). Title. Some Venue. [VERIFY]"); these go in a separate "pending" section of the bib file or are tagged with note = {pending verification}.

Verification criteria for an entry to be considered "verified":

• author list matches the published manuscript (not arXiv preprint if the two differ)
• year is the published year (not arXiv submission year), unless the entry is genuinely arXiv-only
• venue + volume + issue + pages all match the published version
• DOI or arXiv ID is correct (verify by webfetch of the publisher page or the arXiv abstract page)
• citekey follows project convention (typically lastname + year + short-tag, e.g. foster2021dad)

When a verified entry corrects an earlier unverified entry, document the correction in the collection log (Step 5). Do NOT silently overwrite.

Step 2: PDF acquisition

Drop the published-manuscript PDF into references/pdf/<citekey>.pdf, where <citekey> matches the BibTeX citekey exactly. Supplementary material goes as <citekey>-supp.pdf.

The PDFs are gitignored (they're typically large; copyright varies; not appropriate for tracked content). The references/pdf/ directory itself should also be gitignored (or the directory tracked but the *.pdf glob ignored, depending on your project's convention -- see the .gitignore template in templates/paper-skeleton/).

If the published PDF is paywalled and only an arXiv preprint is available, use the preprint and explicitly document this in the collection log so a future reader knows to upgrade when access becomes available.

Step 3: Text extraction

Most AI agents cannot consume PDF binaries directly. Generate a layout-preserving text extraction once, then read it whenever a deep dive is needed:

mkdir -p references/pdf/.txt
pdftotext -layout references/pdf/<citekey>.pdf references/pdf/.txt/<citekey>.txt

pdftotext ships with poppler-utils (Linux: apt install poppler-utils; macOS: brew install poppler). The -layout flag preserves the two-column structure of typical academic papers, which makes the output much more readable than the default flow-mode extraction.

For supplementary material, repeat with <citekey>-supp.pdf -> .txt/<citekey>-supp.txt.

The .txt/ directory is gitignored; it is a regenerable cache.

Batch conversion of all PDFs in one go:

mkdir -p references/pdf/.txt
for pdf in references/pdf/*.pdf; do
  base=$(basename "$pdf" .pdf)
  pdftotext -layout "$pdf" "references/pdf/.txt/${base}.txt"
done

Step 4: Per-paper survey note

Read the .txt extraction, then write notes/survey_<citekey>.md to ~30-50 lines covering the structure in references/survey-note-template.md (loaded on demand from this skill's references/ subfolder when needed).

The survey note is the primary artefact of this workflow: it is what future-you reads when drafting the paper's related-work section, and it is what a collaborator reads to get up to speed without re-reading every PDF.

Briefly, a good survey note has these sections:

1. Header: citekey, full citation, PDF path.
2. Headline claim (one sentence).
3. Method (full detail with key equations in MathJax).
4. Test cases + parameters used in the paper.
5. Headline numerical results.
6. Relevance to our paper (why we cite this; in which sections).
7. Critical observations (insights NOT in the abstract).
8. Action items for our paper draft.

Load references/survey-note-template.md from this skill for the full template + a worked example.

Step 5: Collection log

Append a row to references/_collection_log.md documenting:

• citekey
• verification status (verified / arXiv-only / placeholder)
• date verified
• any corrections made vs an earlier entry
• any quirks worth noting (e.g. "supplementary material is essential", "v3 of the arXiv preprint is significantly different from v1")

The collection log serves as the audit trail. When a reviewer asks "how did you decide on this citation?" or when you discover a mis-attribution months later, the log is where you check.

Load references/collection-log-template.md from this skill for the template.

Workflow rules

1. Never cite a paper you have not verified. Placeholder entries are acceptable in references/bibliography.bib only if they are tagged explicitly (e.g. note = {pending verification}) and marked in the collection log as such.
2. Survey-note quality matters more than count. A short paper with 5 well-written survey notes is more useful than a long one with 25 superficial ones. If a note would just restate the abstract, skip it and link the abstract instead.
3. Re-read the source. When writing the paper's related-work section, re-open the survey note (and if needed the .txt extraction) rather than relying on memory or your initial summary.
4. Track corrections. If you discover a mis-attribution (wrong author list, wrong year, wrong arXiv ID), log it explicitly in _collection_log.md. Do NOT silently fix it -- the audit trail is what makes the corpus trustworthy.
5. MathJax for equations. Survey notes use MathJax ($...$ inline, $$...$$ display) for any equations transcribed from the source. ASCII-art math is forbidden in survey notes.
6. Survey notes and the collection log are HUMAN-FACING documents. Both are read by the user when drafting the paper, by co-authors when re-orienting, and (sometimes) by reviewers. They follow the conventions in the human-facing-doc-authoring skill (audience split, narrative prose over telegraphic fragments, tables where they aid scanning, date-stamps, no personal-path leaks). When producing or substantially revising either artefact, also load ~/.scicomp-research-skills/skills/human-facing-doc-authoring/SKILL.md for the universal conventions and the per-doc-type self-review checklist.

Output contract

When the user invokes this skill, the agent should:

1. Confirm or create the directory layout (Section "Directory layout") if it does not exist.
2. Walk through the 5 steps for each paper the user provides (interactively or in batch).
3. After processing, summarise in a single message:
   • Citekeys added.
   • Verification status of each.
   • Any corrections discovered vs prior entries.
   • PDFs missing (if any).
   • Survey notes written.
   • Suggested next step (typically: read the survey notes, then update the paper's plan-of-record / Section 1 reading list).

See also

• references/survey-note-template.md -- the canonical survey-note template + worked example.
• references/collection-log-template.md -- the canonical collection-log template + worked example.
• templates/paper-skeleton/ (in this repository, sibling of skills/) -- a starter paper-repo skeleton that pre-creates the references/ and notes/ directory structure.
• ~/.scicomp-research-skills/skills/human-facing-doc-authoring/SKILL.md -- universal conventions for human-facing docs. The survey-note + collection-log templates above embody these conventions; load the human-facing-doc-authoring skill when authoring or revising either artefact and you want the cross-cutting checklist.
• ~/.scicomp-research-skills/skills/agent-resource-discipline/SKILL.md -- a literature-survey pass typically processes many PDFs and iterates across sessions, so it is one of the heaviest token consumers in this ecosystem. Load this skill (and in particular its references/pdf-lifecycle.md and references/persistent-memory.md) at the start of any non-trivial literature pass: it codifies the one-shot pdftotext + survey-note-first lookup pattern, the context-window budget for handling many references, and the first-/ last-action index-update protocol that makes a multi-session pass cheap.

Created 2026-05-13 by A. Attia. Distilled from the literature-survey workflow developed for the rl-oed paper (14 references verified across 3 sections of the paper's plan-of-record). Revised 2026-05-13 (added Workflow rule #6 + See-also pointer to human-facing-doc-authoring skill, since survey notes and the collection log are human-facing artefacts). Revised 2026-05-13 (added See-also pointer to agent-resource-discipline skill, since a literature-survey pass is PDF-heavy + multi-session and therefore benefits explicitly from the PDF-lifecycle + persistent-memory + context-window-budget protocols that skill codifies).