Claude-Research-Paper-OS
AGENTS.md4 starsChanged 7 months agoApache-2.0View on GitHub
Claude Code agents are markdown files in .claude/agents/. Each one defines a specialized role. When you ask Claude something that matches an agent's description, Claude spawns it as a subagent with its own context window and tool access. You can also invoke agents explicitly by name. File: .claude/agents/paper-writer.md Triggers when you say: "write," "draft," "rewrite," "improve," or "fix" any section What it does: Rules it enforces (non-negotiable): Output: A LaTeX fragment that fits into main.tex. No \documentclass, no preamble. File:…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 32,671 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository6 other files it gives its agents.
- .claude/CLAUDE.mdCLAUDE.md
- .claude/skills/claude-scientific-writer/CLAUDE.mdCLAUDE.md
- .claude/skills/creative-writing-skills/CLAUDE.mdCLAUDE.md
- content-research-writerSkill
- humanizerSkill
- paper-writerSkill
The file
We split this AGENTS.md into the rules it gives an agent
7 rules. About 0 are common, 7 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 37
No colons or semicolons in prose
- Distinctive: line 38
No AI vocabulary (25+ banned words)
- Distinctive: line 40
Every technical term explained in plain English before first use
- Distinctive: line 41
Every claim hedged with evidence-based language
- Distinctive: line 42
Every formula has plain-English intro + "where" clause
- Distinctive: line 43
Every number attributed to a source
- Distinctive: line 69
If a number is mentioned, where does it come from?
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Agents Reference
Claude Code agents are markdown files in .claude/agents/. Each one defines a specialized role. When you ask Claude something that matches an agent's description, Claude spawns it as a subagent with its own context window and tool access.
You can also invoke agents explicitly by name.
@paper-writer Write the methodology section
@paper-reviewer Review section 3
@citation-checker Check all citations
@humanizer-check Run AI pattern scan
@latex-compiler Compile and check the paper
@deep-researcher Find papers on transformer-based TTSpaper-writer
File: .claude/agents/paper-writer.md
Triggers when you say: "write," "draft," "rewrite," "improve," or "fix" any section
What it does:
- Reads section-specific tips from
agent-research-skills/skills/paper-writing-section/ - Reads your project's
WRITING_RULES.mdif it exists - Reads the existing section in
main.tex - Writes or rewrites the section following all mandatory rules
- Runs the humanizer skill to check for AI patterns
- Runs the writing quality check
- Verifies no colons or semicolons in prose
- Optionally compiles to verify no LaTeX errors
Rules it enforces (non-negotiable):
- No colons or semicolons in prose
- No AI vocabulary (25+ banned words)
- No em dashes
- Every technical term explained in plain English before first use
- Every claim hedged with evidence-based language
- Every formula has plain-English intro + "where" clause
- Every number attributed to a source
- Same tone from start to end
- Sentences connect naturally
- Paragraph and sentence length varies
Output: A LaTeX fragment that fits into main.tex. No \documentclass, no preamble.
paper-reviewer
File: .claude/agents/paper-reviewer.md
Triggers when you say: "review," "check," "audit," or "verify"
What it does:
This is the most thorough agent. It runs a word-by-word devil's advocate review across 13 quality dimensions.
The Devil's Advocate Protocol:
For every line of prose, it asks these questions.
Term-level:
- Is this technical term explained before it is used?
- Would a school student understand this word?
- If a number is mentioned, where does it come from?
Sentence-level:
- Does this sentence connect to the previous one?
- Is there a jump between ideas?
- Is this claim stated as absolute fact or properly hedged?
- Can I remove this sentence without losing meaning?
Paragraph-level:
- Does this paragraph have one clear purpose?
- Does the first sentence tell me what the paragraph is about?
- Does the last sentence lead into the next paragraph?
Section-level:
- Does this section tell a story from beginning to end?
- Is the tone consistent with the abstract and introduction?
- Does every table/figure appear near its first reference?
The 13 Review Dimensions:
- Prose quality (writing quality checklist)
- AI pattern detection (25 flagged terms, em dashes, synonym cycling)
- Colon/semicolon check (banned in prose)
- Academic humility (every claim hedged)
- Formula completeness (full "where" clause)
- Citation accuracy (automated validation script)
- Figure/table placement (near first reference)
- Flow and tone (consistent voice throughout)
- Repetition (key statistics not overused)
- Research accuracy (claims match cited sources)
- Term explanation (every technical term explained)
- Number attribution (every number has a source)
- Sentence connection (every sentence links to the previous)
Output: A structured report with exact line numbers, problematic text, severity level (Critical/Important/Minor), the devil's advocate question that caught it, and a suggested fix.
citation-checker
File: .claude/agents/citation-checker.md
Triggers when you say: "check citations," "verify references," or "validate bibliography"
What it does:
- Runs the automated citation validation script
python ~/.claude/skills/agent-research-skills/skills/citation-management/scripts/validate_citations.py --tex main.tex --bib references.bib - For each key numerical claim, verifies the cited source matches
- Scans all prose for factual/quantitative statements with no
\cite{}nearby - Checks for stale references (broken arXiv IDs, missing authors)
Output: A report with three sections.
- Citation validation results (from script)
- Uncited claims (line number, text, suggested citation)
- Reference quality issues (incomplete entries, missing venues)
humanizer-check
File: .claude/agents/humanizer-check.md
Triggers when you say: "humanize," "check AI patterns," "make it sound human," or after any writing session
What it does:
Scans the entire paper (or a specific section) for signs of AI-generated text across 5 check categories.
A. Flagged Terms (25+ words) Scans for high-frequency AI vocabulary. For each found, suggests a replacement.
B. Punctuation Patterns
- Em dashes in prose (must be 0)
- Semicolons in prose (must be 0)
- Colons in prose (must be 0)
C. Structural Patterns
- Rule-of-three forcing (every list has exactly 3 items)
- Uniform paragraph length (all 150-200 words)
- Synonym cycling (calling the same thing by different names)
- Mirror structure (every section has identical internal format)
D. Tone Patterns
- Throat-clearing openers ("In the realm of")
- Meta-commentary ("This section will discuss")
- Significance inflation ("stands as," "serves as")
- Copula avoidance ("represents a" where "is a" works)
E. Sentence Rhythm (Burstiness) Flags any 5+ consecutive sentences with similar word count.
Output: For each issue found, reports line number, current text, pattern matched, and the fix.
latex-compiler
File: .claude/agents/latex-compiler.md
Triggers when you say: "compile," "build," or "check the paper"
What it does:
- Compiles the paper with Docker (or local pdflatex)
docker run --rm -v "$(pwd)":/data blang/latex:ubuntu bash -c \ "cd /data && pdflatex -interaction=nonstopmode main.tex && \ bibtex main && pdflatex -interaction=nonstopmode main.tex && \ pdflatex -interaction=nonstopmode main.tex" - Verifies "Output written on main.pdf" appears
- Checks page count
- Checks for warnings (undefined references, missing citations)
- Runs the LaTeX quality checker
python ~/.claude/skills/agent-research-skills/skills/latex-formatting/scripts/latex_checker.py main.tex - Runs the assembly checker
python ~/.claude/skills/agent-research-skills/skills/paper-assembly/scripts/assembly_checker.py --dir . --verbose
Output: Compilation status, page count, any warnings or errors, and quality check results.
deep-researcher
File: .claude/agents/deep-researcher.md
Triggers when you say: "research," "find papers," "verify this claim," or "check novelty"
What it does:
Searches three academic databases in parallel.
# Semantic Scholar
python ~/.claude/skills/agent-research-skills/skills/deep-research/scripts/search_semantic_scholar.py \
--query "QUERY" --max-results 20
# CrossRef
python ~/.claude/skills/agent-research-skills/skills/literature-search/scripts/search_crossref.py \
--query "QUERY" --rows 10
# arXiv source download
python ~/.claude/skills/agent-research-skills/skills/literature-search/scripts/download_arxiv_source.py \
--arxiv-id "ARXIV_ID"Four research modes:
- Claim Verification -- Given a claim and citation, verify they match
- Gap Analysis -- Search for work that might fill a gap you identified
- Novelty Check -- Search for prior surveys that might overlap with yours
- Update Check -- Search for new papers published since your last review
Output: Structured results with title, authors, year, venue, and relevance assessment for each paper found.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

