academic-research-harness

name	academic-research-harness
description	Use when explicitly requested for paper-writing or literature-grounded work: paper search/indexing, BibTeX/reading notes, idea refinement, repo-to-paper Markdown drafting, claim/citation audits, writing-style revision, or paper handoff.

Academic Research Harness

Purpose

Use this skill as the paper-writing and evidence layer for literature, indexes, claims, Markdown sections, citation audits, and handoff.

Keep code, training, debugging, validation, and experiment execution outside unless the user explicitly connects them to paper writing.

Activation Boundary

Use only for explicit paper-writing or literature-grounded work. Coding, debugging, training, logs, and general research stay outside unless tied to papers, claims, citations, or sections.

Progressive Disclosure Rule

Start here, choose the route, then load only needed references and templates.

Create workspaces progressively. Use full only on request; otherwise create the route's minimal files and add section templates only when drafting them.

Task Router

User intent	Load	Workspace mode	Template loading
Initialize or validate a paper workspace	references/workspace.md	requested mode, default minimal	only files being created or checked
Search, collect, index, download, classify, or summarize papers	references/literature.md	literature	paper_index.md, references.bib, notes/<citation_key>.md, and papers/*.pdf only when writing/downloading them
Judge source freshness, venue priority, arXiv credibility, or downgrade/exclusion	references/source-quality.md	literature or citation-audit	paper_index.md or audited claim evidence only
Refine an idea using papers	references/literature.md	idea	idea_log.md only when writing it
Convert repo/code/config/logged setup into paper sections	references/repo-to-paper.md	repo-to-paper --section <target>	only the requested section template plus claims.md when adding claims
Create result tables or placeholders	references/repo-to-paper.md	repo-to-paper --section results_tables	results_tables.md
Draft or revise writing style, contribution framing, limitation placement, or over-defensive prose	references/writing-style.md	existing mode	target section plus claims.md when claim handling changes
Audit citations, claims, novelty, related work, or SOTA language	references/citation-audit.md	citation-audit	claims.md, audited section, and only the needed evidence files (paper_index.md, references.bib, relevant notes/sources)
Prepare paper-state handoff	references/handoff.md	handoff	handoff.md only
Check shared evidence, results, or confirmation constraints	references/evidence-policy.md	existing mode	only files needed to support the decision

Shared Rules

• For workspace creation, validation, mode selection, and scripts, load references/workspace.md.
• For source freshness, venue priority, arXiv credibility, or downgrade/exclusion, load references/source-quality.md.
• For novelty, SOTA, superiority, results, contribution framing, claim support, or confirmation boundaries, load references/evidence-policy.md.
• For writing style, contribution framing, limitation placement, or over-defensive prose, load references/writing-style.md.
• If a task spans routes, load relevant references but keep template loading minimal.
• Use available browsing/search tools with site: filters when supported. Use available fetch/open tools for abstracts/full text and local file tools for repo evidence. Do not invent tool names.
• For cross-file consistency, inspect only the necessary workspace files, normally venue_profile.md, paper_index.md, claims.md, and the target section.
• Keep raw engineering identifiers out of paper prose; translate them through repo-to-paper or writing-style rules.
• Record nontrivial paper claims in claims.md.
• Report results only from user-provided numbers or verified logs.

Route Boundaries

• Literature collection must not automatically draft intro.md, related_work.md, or method.md.
• Idea refinement writes candidates to idea_log.md; it must not modify code or convert suggestions into final claims without evidence and user decision.
• Repo-to-paper writes only the requested section and marks unknown implementation details as TODO or unverified.
• Citation audit checks exact claim support, not topical similarity.
• Handoff updates handoff.md only for explicit handoff or material Tier A/B paper-state changes.