ParaSkills

A scientific-research-workflow skill library that turns five specialist 4B models into many task-specific research subagents. Each skill is an Anthropic Agent Skills / agentskills.io SKILL.md — portable across Claude Code, Cursor, Codex, and any agent that loads the open format.

ParaSkill = scientific research workflow skill
subagent  = base_model + skill + role profile + input scope + output contract

Skill 不会凭空产生新能力——能力上限来自 SFT 和基础模型。Skill 把能力变得可控、可复用、可编排。

5 个专家模型 × 28 个 skill

围绕一篇顶会论文的闭环:Creator(产 idea) → Designer(规划实验) → Writer(写) → Reviewer(审) → Checker(查闭环)。

family	count	代表 skill
writer	8	revise-by-criterion, amplify-genuine-strengths, reposition-related-work, expand-analytical-discussion, reframe-limitations-honestly, diagnose-aha-moment, rewrite-abstract-with-commitments, build-intro-argument-chain
reviewer	6	simulate-top-conference-review, identify-fatal-concerns, calibrate-rating-severity, judge-review-shift, profile-reviewer-signals, judge-revision-winrate
checker	6	check-claim-evidence-alignment, check-section-logic-consistency, detect-overclaim-missing-closure, detect-presentation-gaming, detect-limitation-laundering, enforce-content-preservation-zones
creator	4	reverse-engineer-paper-idea, interactively-refine-idea, generate-idea-variants, position-novel-idea
designer	4	map-claims-to-experiments, design-baselines-and-ablations, design-diagnostic-analysis, design-defensive-experiments

完整路由表(含 trigger / required_inputs / optional_inputs / handoff_to)见 SKILL_INDEX.json。

每个 SKILL.md 的结构

YAML frontmatter(必填 name + description,扩展放 metadata:) + 正文 10 段:Purpose / When to Use / Do Not Use / Required Inputs / Output Contract / Procedure / Quality Bar / Failure Modes / Handoff / Example。

来源 / grounding

• 原始 15 个(5 family × 3)为本项目设计。
• 另外 13 个提炼自两篇论文(数据集构建部分未纳入):
  • ARGAR(arXiv 2606.13044, Adversarial Repackaging Gaming AI Review):把攻击知识翻转成 Checker 防御(detect-presentation-gaming / detect-limitation-laundering / enforce-content-preservation-zones),其 strength-weakness asymmetry 与 reframing>polishing 发现支撑 Writer 的新 skill。
  • XtraGPT(arXiv 2505.11336, criteria-guided context-aware revision):revise-by-criterion 是其原生 paper-revision skill 的改进移植;Length-Controlled Win Rate → judge-revision-winrate。

使用 / 校验

# 校验全部 skill 并(重)生成路由表
python tools/check_paraskills.py            # 校验 + 生成 SKILL_INDEX.json
python tools/check_paraskills.py --check    # 只校验

checker 会校验:frontmatter 必填字段、正文 10 段、family/model 一致性、optional_inputs、以及 handoff_to 是否指向真实存在的 skill(路由图闭合)。

新增 skill:复制 TEMPLATE_SKILL.md 到 <family>/<skill-name>/SKILL.md,填好再跑 checker。

License

MIT。两个参考项目(ARGAR、XtraGPT)亦为 MIT。