Paper Discourse Graph

Paper Discourse Graph

Use this skill to review a paper section as a reader-state graph, not only as a linear outline. The goal is to see what each paragraph plants, what it pays off, where notation appears before intuition, and where a justification detour interrupts the main story.

1. Workflow

1.1 Establish scope

Identify the source .tex file or section to audit. If the user wants discussion before editing, produce findings and proposed interventions only.

Look for project-local guidance before running a generic pass:

• AGENT_REVIEW_BRIEF.md
• docs/*review*brief*.md
• tools/discourse_graph/profiles/*.json
• paper-specific profiles or terminology docs

1.2 Choose a profile

Use the most specific available profile:

1. user-provided --profile
2. project-local profile, if present
3. bundled references/profiles/cphast.json for C-PHAST drafts
4. bundled references/profiles/default_scientific_paper.json

Profiles define the central reader question, audience needs, domain terms, and topic checks. Keep project taste in profiles, not in the engine.

For methods, theory, or appendix sections where paragraphs hand off to notation and equations, also read references/formal_block_flow.md. When authoring or revising source files and the user wants durable machine-readable breadcrumbs for later passes, read references/source_semantic_comments.md. For late-stage paragraph-by-paragraph refinement, task breakdowns for fresh agents, or requests to zoom into each paragraph with a fresh mind, read references/paragraph_refinement_tasks.md.

1.3 Run the CLI

From this skill directory:

python scripts/discourse_graph_audit.py  \
  --section  \
  --profile references/profiles/default_scientific_paper.json \
  --out  \
  --json-out 

For C-PHAST:

python scripts/discourse_graph_audit.py paper/sec/3_method_arxiv.tex \
  --section 3 \
  --profile references/profiles/cphast.json \
  --out paper/docs/plans/discourse_graph_sec3.md \
  --json-out paper/docs/plans/discourse_graph_sec3.json

Optional LLM chunks:

python scripts/discourse_graph_audit.py  \
  --section  \
  --out  \
  --llm-jsonl 

1.4 Interpret the report

Treat the graph as triage, not truth. The labels are prompts for manual review.

Important node roles:

• scene: visible example or running situation
• question: planted reader question
• claim: assertion or section-level promise
• mechanism: method component or construction
• definition / notation: formal object introduction
• justification: reason, proof sketch, or defensive support
• evidence: figure, table, theorem, algorithm, or equation block
• boundary: scope limit or condition
• payoff: answer to earlier setup
• handoff: transition to the next object
• semantic_comment: source-only % DG: breadcrumb used to record

author intent without rendering in the PDF

Important risk labels:

• abrupt: weak continuity into the current block
• detour: useful material that may interrupt the active story
• unpaid_question: a planted question without nearby payoff
• premature_notation: dense formal terms before visible grounding
• weak_parent_link: paragraph is weakly tied to its heading
• bridge_candidate: possible bridge that may rescue a detour
• unconnected_evidence: figure/table/algorithm not clearly recalled
• context_debt: named setting or experiment appears before enough

local setup

• appendix_claim_leak: appendix-only evidence is doing main-text

argumentative work

• symbol_alias_confusion: profile-tracked symbols gain a new

subscript or variant without explaining the relationship

• coarse_algorithm_reference: algorithm behavior is described

without line, stage, equation, or step anchors

• missing_formal_setup: equation-like block appears before prose

creates the need for it

• missing_formal_payoff: equation-like block is not followed by a

nearby interpretation

• symbol_scope_debt: formal block introduces several new symbols

without nearby scope, type, or dependency prose

• role_mismatch: prose promises intuition or readability but hands

off to dense notation without enough setup

• equation_overload: formal block introduces many symbolic objects

at once

• dangling_reference: paragraph starts from this/these/such object

without a clear local antecedent

2. Review Taste

For each paragraph or block, ask:

• What does the reader know right now?
• What question did the previous block plant?
• Does this block answer, sharpen, or defer that question?
• Is the payoff close enough, or is the reader carrying too much load?
• Is this block story-bearing, justification-bearing, evidence-bearing,

implementation-bearing, or boundary-bearing?

• If this became a task for a fresh agent, what local context,

invariants, and claim boundaries would that task need?

Prefer a chain where understanding progressively accumulates: visible scene -> reader question -> mechanism -> formal object -> evidence -> payoff -> next question.

Use visible language before formal terms. A dense equation should feel inevitable because the previous paragraph made the need for that object obvious.

For formal blocks, check the local loop: paragraph motivates need -> notation names object -> equation commits to definition/update/loss/constraint/readout -> paragraph interprets what changed for the reader.

3. Output Format

Report findings with:

• file and line range
• live reader question
• risk label
• why the current placement breaks or helps the story
• smallest proposed intervention

Use labels:

• Payoff: answers a live reader question
• Plant: intentionally creates the next question
• Bridge: connects intuition to a formal object
• Detour: useful but not needed for the current story step
• Stray: does not support the current claim chain

When the user asks to discuss, do not edit. Rank the highest-value changes and ask which ones to land.

4. Bundled Resources

• scripts/discourse_graph_audit.py: command-line entry point
• scripts/discourse_graph/: dependency-free LaTeX segmentation,

heuristics, schema, and rendering code

• references/formal_block_flow.md: detailed rubric for paragraphs,

notation, equations, interpretation, and consistency

• references/paragraph_refinement_tasks.md: late-stage per-paragraph

task protocol for fresh review passes

• references/source_semantic_comments.md: optional % DG: comment

convention for source-level discourse breadcrumbs

• references/profiles/default_scientific_paper.json: general profile
• references/profiles/cphast.json: C-PHAST profile with typed

factors, charts, ports, PHASTCore, and action-card checks

5. Pitfalls

• Do not treat lexical overlap as a correctness judgment.
• Do not force every paragraph to use explicit transition phrases.
• Do not remove rigorous caveats merely because they are detours;

decide whether they belong later, in an appendix, or in a tighter boundary sentence.

• Do not let profile-specific vocabulary leak into unrelated papers.
• Keep paper-specific notation families in the profile, for example

risk_rules.confusable_symbol_bases, not in the reusable engine.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: shubham0704
• Source: shubham0704/claude-skills
• License: MIT
• Homepage: https://github.com/shubham0704/claude-skills

Install and usage instructions live in the source repository linked above.