Academic Paper Writer — Claude Agent Skill

A production-grade Claude Agent Skill for international-standard (Q1-level) Persian (فارسی) academic writing, with general coverage across scientific fields and extra depth in security & networking.

The skill is intentionally thin at the top and heavy in references (Progressive Disclosure): SKILL.md routes each task to the right reference file; the references hold the deep domain knowledge that is loaded only when needed.

What it does

Three operating modes:

1. Full paper from scratch — abstract → references, as a responsible framework with [نیازمند ارجاع] / [نیازمند داده] placeholders (never fabricated findings).
2. Section-by-section — مقدمه / روش / یافته‌ها / بحث / نتیجه‌گیری / چکیده, each with its own section logic (CARS for the intro, reproducibility for methods, "so what?" for discussion).
3. Edit / upgrade existing text — formalize register, fix overclaiming, balance hedging, surface citation/data gaps.

It also covers submission prep (IEEE / APA7 / Vancouver), citation audits, EQUATOR checklists, and polite point-by-point reviewer responses.

Structure

academic-paper-writer/
├── SKILL.md                              # Thin router: frontmatter + Quick Reference + protocol
├── references/
│   ├── structure-templates.md            # Article types, IMRaD, CARS, EQUATOR, abstracts
│   ├── persian-academic-style.md         # Register, hedging, tense, before/after examples
│   ├── citation-and-sourcing.md          # Integrity rule, APA7/IEEE/Vancouver, CRediT/ICMJE, AI policy
│   ├── security-networking-domain.md     # Safe framing, threat model, STRIDE/PASTA, security metrics
│   └── common-mistakes.md                # 7-axis rubric, stats mistakes, R.E.S.P.O.N.D./RARRR responses
└── evals/
    ├── evals.json                        # 7 realistic, discriminating capability evals
    └── trigger-eval.json                 # 20-query trigger audit (10 should-trigger / 10 near-miss)

Design principles (per the skill-writing standard)

• Progressive Disclosure — metadata always in context; SKILL.md body (~95 lines) on trigger; references on demand.
• WHY over MUST — rules explain their reasoning (e.g. why not to fabricate citations) instead of bare imperatives.
• Lean — no content Claude already knows; every line earns its place.
• Safety — security topics stay scholarly, defensive, and non-operational (no exploit code, scanning, or evasion steps).
• No fabrication — facts, results, statistics, and references are never invented; gaps are marked with placeholders.

Testing surface

• evals/evals.json — 7 capability evals with discriminating expectations (a hallucinated or overclaimed output fails).
• evals/trigger-eval.json — 20 realistic queries auditing description triggering, including hard near-misses (pure translation, marketing whitepaper, slides, English paper, port-scanning script).

License

MIT — see LICENSE.