claims-map

A Claude Code skill that maps a paper's claims across its sections. For every claim made in the abstract, introduction, conclusion and contributions list, it shows what each body section says about it, with a quoted anchor: anchored, partial, differs, absent, or external, unverified (support is a citation).

It checks consistency, not correctness. It never says a claim is right or wrong, never proposes wording, and never reads cited papers. The author decides the story; the map shows where the text agrees with itself.

Status: experimental. Piloted on two LaTeX papers (one a CC BY arXiv paper) with Sonnet 5.5. Expect rough edges; see CHANGELOG.md and the known limits below.

This skill and its scripts were written with Claude (Anthropic's AI model) in Claude Code. The agents it runs are LLM agents too, which is why every verdict must carry a quote you can check.

Example

This is the real output for the toy paper in examples/toy-paper/. The briefing, briefing.md:

# Briefing: Toy paper

6 claims in Toy paper: 2 established, 1 conflicting, 1 narrower, 1 cited, 1
unsupported.

- **T06, unsupported.** "This paper learns the mapping instead." Claimed in
  introduction, sentence 2; no body section states it.
- **T05, conflicting.** "We train the classifier in three stages." (abstract,
  sentence 4). A body section differs. results: "Phase 2 is evaluated second."
- **T03, narrower.** "It beats the baseline." (abstract, sentence 3). The body
  states a narrower version. results: "The method is twice as fast as the older
  design."
- **T04, cited.** "Prior work uses Kalman filters \cite{smith2020}." (intro,
  sentence 1). The support is a cited work, which was not read.
- **T01 and T02, established.** A body section shows each of them; the quotes
  are in the ledger.

Nothing here says whether any claim is true. It shows where the text agrees with
itself.

The top of the claim ledger, claim_ledger.md, worst first:

id	status	claim	where claimed	evidence
T06	Unsupported	This paper learns the mapping instead.	introduction, sentence 2	none: absent from methods and results
T05	Conflicting	We train the classifier in three stages.	abstract, sentence 4	results differs: "Phase 2 is evaluated second."
T03	Narrower	It beats the baseline.	abstract, sentence 3	results partial: "The method is twice as fast as the older design."

Each claim gets one status: Conflicting, Established, Assumed, Narrower, Cited or Unsupported. Every verdict is backed by a quote, and the ledger's detail section gives the file and line it came from. Established means the paper shows the claim, not that the claim is true. Nothing here is advice: the map reports where the text agrees with itself and never says what to change.

Install

Clone into your Claude Code skills folder:

git clone https://github.com/theMuslimDev/claims-map.git ~/.claude/skills/claims-map
pip install pyyaml

Copy claims-map.config.template.md next to your paper, fill in the YAML block, then ask Claude Code for a claims map of your paper (the skill triggers on that wording). Sub-agents default to Sonnet 5.5; set agent_model in the config to override. Next steps are in USAGE.md.

As a plugin (experimental)

claude plugin marketplace add theMuslimDev/claims-map
claude plugin install claims-map@claims-map-marketplace

The repo root is the plugin, so a plugin install and a manual clone give the same files.

Versions

Installs from the plugin marketplace follow the pinned release (the ref in .claude-plugin/marketplace.json) and stay on it until the next release ships. Cloning the repository follows main, which moves ahead of the last release. Releases are listed in CHANGELOG.md.

Quick start

After installing, ask Claude Code for a claims map:

/claims-map:claims-map /path/to/paper.pdf

Give it a PDF, DOCX, Markdown or text file, or a LaTeX folder. Claude drafts a config, you confirm the sections, and it writes briefing.md and claim_ledger.md (plus claims_map.md, and an optional HTML matrix) to the output folder, leaving your files untouched.

Full walkthrough: USAGE.md.

How it works

1. Snapshot the section files (sha256, git commit, dirty-tree flag). All agents read the snapshot, never the live tree.
2. Extractor agent writes a claims table from the front matter. It gathers no evidence.
3. One section agent per body section, in parallel. Each sees only the claims table and its own section.
4. Number check (a script) compares numbers for the same quantity across sections: equal, rounding, conflict.
5. Build the markdown map (the source of truth), enforcing the anchor rule mechanically, plus a terminology lint from your config. From the same checked cells it writes the claim ledger (claim_ledger.md, one status per claim, worst first) and a plain-text briefing (briefing.md), the main outputs.
6. Render a self-contained HTML matrix (optional).

Hard rules: no anchor, no verdict; agents report and never fix; section agents are independent; terminology and framing rules come from your per-paper config, not the skill.

Known limits

• PDF input is new: text extraction, section splitting and the config were tested on a toy PDF and one real IEEE-style paper, but the full agent workflow has not been run on a PDF yet.
• Consistency only: two sections can agree and both be wrong.
• The anchor check proves a quote exists; a weak-anchor flag warns when claim and quote share no content word, but relevance is still a judgement.
• The number check is a heuristic and can pair unrelated numbers.
• Figures, tables and captions are visible only if listed as sections or support files.
• Best on LaTeX. PDF input depends on text extraction: two-column layouts, equations, tables and figures can come out garbled, and anchors point at the extracted text, not PDF pages.

Tests

python -m unittest discover -s tests

94 tests, none needing a model: an end-to-end run on examples/toy-paper/ (whose golden output is the Example above), plus unit tests. A manual acceptance test on a real CC BY arXiv paper is in examples/arxiv-2610.24343/ (needs model agents; see its README).

Licence

See LICENSE.