zxyasfas/paper_format_agent

Paper Format Agent

Use this skill for local-first academic document formatting work. It is designed for .docx papers and a separate format specification in .docx, .doc, or .txt.

Safety Defaults

• Never upload private papers or format guides to external services unless the user explicitly asks for it.
• Treat student papers, manuscripts, reviewer comments, and school templates as sensitive data.
• Do not change academic content unless the user explicitly opts in. Default to formatting-only changes.
• Keep reports and generated documents in a user-provided output directory.

Standard Workflow

1. Confirm the input files:
   • --format-file: format guide or template text.
   • --paper-file: source paper, currently .docx.
   • --out-dir: output directory for reports and generated documents.
2. Run the CLI in formatting-only mode:

python -m paper_format_agent.cli \
  --format-file "format_guide.docx" \
  --paper-file "paper.docx" \
  --out-dir "./output" \
  --engine auto \
  --strict-required-sections

3. Inspect generated artifacts:
   • formatted_paper_v3.docx: repaired document.
   • format_rules.json: extracted rules.
   • format_report.json: machine-readable score and checks.
   • format_report.html: human-readable report.
   • modify_log.json: formatting operations.
   • engine_report.json: post-processing engine result.
4. Check the content guard fields in format_report.json:
   • content_changed should normally be false.
   • content_guard_enforced should normally be true.
5. If a rule is wrong, update rule extraction or scoring logic and add a minimal test case.

Validation

Run these before handing work back:

python tools/validate_skill.py
python -m unittest discover -s tests -p "test_*.py"
python tools/compile_check.py
python tools/release_audit.py

When Adding Template Support

• Add the smallest representative rule text needed to reproduce the behavior.
• Prefer deterministic rule extraction over LLM-only interpretation.
• Add tests for extracted margins, font size, line spacing, required sections, headings, captions, or references.
• Do not commit real student documents. Use synthetic text or anonymized fixtures.

When Reviewing Output

Read format_report.json first, then inspect format_report.html if the user needs a human-readable summary. Prioritize issues in this order:

1. Content changed unexpectedly.
2. Required sections or headings were misclassified.
3. Page setup, margins, fonts, or line spacing are wrong.
4. Captions, references, tables, or numbering look wrong.
5. Report wording or score explanation is unclear.

Contribution-Friendly Tasks

Good first PRs usually fit one of these buckets:

• Add a synthetic test for a school, journal, or conference formatting rule.
• Improve a rule extractor with a narrowly scoped regex or parser.
• Improve report clarity without changing scoring semantics.
• Add a regression case to docs/regression_manifest.sample.json.
• Improve this skill workflow or agents/openai.yaml metadata.