search-prior-art
について
このスキルは、特許、学術研究、製品、オープンソースソフトウェアを横断して構造化された先行技術調査を行い、発明の新規性を評価します。開発者は、特許性の評価、既存特許への異議申し立て、または実施自由権(FTO)分析の支援に活用できます。関連する既存の開示情報を見つけるために、ウェブ検索とデータ取得ツールを活用します。
クイックインストール
Claude Code
推奨npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/search-prior-artこのコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします
ドキュメント
Search Prior Art
Structured prior art search → find pubs|patents|products|disclosures predating invention. Used for patentability (can patent?), validity challenge (should have been granted?), FTO (covered by existing rights?).
Use When
- Eval novelty+non-obvious pre-file
- Challenge existing patent validity → find art examiner missed
- Support FTO → find art limiting blocking patent scope
- Document defensive pub → prevent others patenting concept
- Respond to office action questioning novelty|obviousness
In
- Required: Invention desc (what, how, problem)
- Required: Purpose (patentability|invalidity|FTO|defensive)
- Required: Critical date (filing date or invention date)
- Optional: Known related patents|pubs
- Optional: Tech classification codes (IPC, CPC)
- Optional: Key inventors|companies
Do
Step 1: Decompose Invention
Break into constituent technical features.
- Read desc (or claims if vs existing patent)
- Extract essential elements — each independent feature:
- Components?
- Process steps?
- Technical effect?
- Problem + how solved?
- ID novel combination — what's diff from known:
- New element added to known?
- New combo of known?
- Known element new field?
- Gen search terms per element:
- Tech terms, synonyms, abbrev
- Broader+narrower (hierarchy)
- Alt descriptions
- Doc Search Map: elements, terms, relationships
Search Map Example:
+------------------+-----------------------------------+-----------+
| Element | Search Terms | Priority |
+------------------+-----------------------------------+-----------+
| Attention layer | attention mechanism, self- | High |
| | attention, multi-head attention | |
| Sparse routing | mixture of experts, sparse MoE, | High |
| | top-k routing, expert selection | |
| Training method | knowledge distillation, teacher- | Medium |
| | student, progressive training | |
+------------------+-----------------------------------+-----------+
→ Complete decomposition w/ terms per element. Novel combo ID'd → search must find (invalidate) or confirm absent (support novelty).
If err: too abstract → ask more specific. Claims unclear → broadest reasonable interp per element.
Step 2: Search Patent Literature
Patent DBs systematic.
- Construct queries:
- Each element individually first (broad)
- Combine to find closer art (narrow)
- Classification codes filter by tech area
- Multi DBs:
- Google Patents: Full-text, free, large
- USPTO PatFT/AppFT: US patents+apps, official
- Espacenet: EU, excellent classification
- WIPO Patentscope: PCT, global
- Date filters:
- Prior art must predate critical date
- Up to 1yr pre-filing (grace varies by jurisdiction)
- Per relevant result record:
- Doc number, title, filing date, pub date
- Which elements disclosed (map to Search Map)
- Discloses novel combo?
- Classify by relevance:
- X: Discloses invention alone (anticipation)
- Y: Key elements, combinable (obviousness)
- A: Background art
→ Classified patent ref list mapped to elements. X (if found) = showstoppers for novelty. Y = building blocks for obviousness.
If err: no relevant patent art → doesn't mean novel — non-patent (Step 3) may have critical ref. Absence in 1 DB ≠ absence everywhere.
Step 3: Non-Patent Literature
Academic, products, OSS, other.
- Academic:
- Google Scholar, arXiv, IEEE Xplore, ACM Digital Library
- Same Step 1 terms
- Conf papers + workshop proceedings often predate patents
- Products + commercial:
- Product docs, manuals, marketing
- Internet Archive (Wayback) for date-verified web
- Trade pubs + press releases
- OSS + code:
- GitHub, GitLab — search impls of features
- READMEs, docs, commit history for date evidence
- Software releases w/ ver dates
- Standards:
- IEEE, IETF (RFCs), W3C, ISO
- Standards-essential patents must be disclosed; search standard bodies' IP DBs
- Defensive pubs:
- IBM Technical Disclosure Bulletin
- Research Disclosure journal
- IP.com Prior Art DB
- Verify pub date before critical date:
- Web: Wayback for date evidence
- Software: release dates|commit timestamps
- Papers: pub date not submission
→ Non-patent refs complement patent search. Academic + OSS often most powerful — describe details more explicitly than patents.
If err: non-patent sparse → tech primarily corp R&D (patent-heavy). Shift emphasis to patent + combo-based obviousness.
Step 4: Analyze + Map
Eval how art relates to invention.
- Claim chart mapping art → elements:
Claim Element vs. Prior Art Matrix:
+------------------+--------+--------+--------+--------+
| Element | Ref #1 | Ref #2 | Ref #3 | Ref #4 |
+------------------+--------+--------+--------+--------+
| Element A | X | X | | X |
| Element B | | X | X | |
| Element C | X | | X | |
| Novel combo A+B+C| | | | |
+------------------+--------+--------+--------+--------+
X = element disclosed in this reference
- Novelty: Single ref discloses all elements?
- Yes → anticipated (not novel)
- No → may be novel (proceed obviousness)
- Obviousness: Few refs (2-3) combinable to cover all?
- Motivation to combine? (skilled person sees reason?)
- Teach away? (suggest wouldn't work?)
- FTO: Does art narrow blocking patent claims?
- Art overlapping blocking patent's claims limits enforceable scope
- Document analysis w/ specific passage citations
→ Clear claim chart showing element coverage by refs, w/ novelty + obviousness assessment. Each mapping cites specific passages|figures.
If err: chart shows gaps (elements not in any art) → those = potentially novel. Focus follow-up on specific gaps.
Step 5: Document + Deliver
Package for intended use.
- Write Prior Art Search Report:
- Purpose + scope
- Methodology (DBs, queries, date ranges)
- Results summary (count, classification breakdown)
- Top refs w/ detailed analysis (claim charts)
- Assessment: novelty, obviousness, FTO implications
- Limitations + further-search recommendations
- Organize refs:
- Sorted by relevance (X first, Y, A)
- Each w/ full bibliographic + access link
- Key passages highlighted|extracted
- Recommendations by purpose:
- Patentability: File|don't, suggested claim scope by gaps
- Invalidity: Strongest combo, suggested legal arg
- FTO: Risk level, design-around opportunities, licensing
- Defensive: Whether to publish defensive disclosure based on white space
→ Complete organized report directly supporting decision. Refs accessible, analysis traceable.
If err: inconclusive (no strong X|Y, but relevant background) → state clearly: "No anticipatory art; closest addresses A+B not C. Recommend file w/ claims emphasizing C." Inconclusive valid + useful.
Check
- Invention decomposed into searchable elements
- Novel combo explicitly ID'd
- Patent DBs searched (min 2)
- Non-patent searched (academic + products + OSS)
- All refs predate critical date (verified)
- Claim chart maps elements w/ passage citations
- Novelty + obviousness assessed w/ reasoning
- Classified (X, Y, A)
- Report has methodology, limitations, recommendations
- Reproducible (queries + DBs documented)
Traps
- Keyword tunnel vision: Exact terms only misses synonyms. Use Step 1 hierarchy.
- Patent-only search: Non-patent (papers, products, code) often more explicit. Don't skip Step 3.
- Date carelessness: Must predate critical date. Brilliant ref 1 day after = worthless.
- Ignore foreign: Major inventions may first appear in CN|JP|KR|DE patents. MT makes searchable.
- Confirmation bias: Searching to confirm novelty vs to find invalidating art. Best search tries hardest to find closest.
- Stop too early: First results rarely best. Iterate based on field vocabulary revealed.
→
assess-ip-landscape— broader landscape mappingscreen-trademark— TM-specific (diff DBs + legal frame than patent)file-trademark— TM filing post-screenreview-research— lit review methodology overlapssecurity-audit-codebase— systematic methodology parallels (thoroughness, doc, reproducibility)
GitHub リポジトリ
関連スキル
executing-plans
デザインexecuting-plansスキルは、完全な実装計画があり、それを管理されたバッチでレビューチェックポイントを設けながら実行する場合に使用します。このスキルは計画を読み込んで批判的にレビューした後、小さなバッチ(デフォルトは3タスク)でタスクを実行し、各バッチの間に進捗状況を報告してアーキテクトのレビューを受けます。これにより、品質管理チェックポイントが組み込まれた体系的な実装が保証されます。
requesting-code-review
デザインこのスキルは、コードレビュアーサブエージェントを起動し、処理を進める前に要件に対してコード変更を分析します。タスク完了後、主要な機能の実装後、またはmainブランチへのマージ前などに使用すべきです。このレビューは、現在の実装と元の計画を比較することで、問題を早期に発見するのに役立ちます。
connect-mcp-server
デザインこのスキルは、開発者がHTTP、stdio、またはSSEトランスポートを使用してMCPサーバーをClaude Codeに接続するための包括的なガイドを提供します。GitHub、Notion、カスタムAPIなどの外部サービスを統合するためのインストール、設定、認証、セキュリティについて解説しています。MCP統合のセットアップ、外部ツールの設定、またはClaudeのModel Context Protocolを扱う際にご利用ください。
web-cli-teleport
デザインこのスキルは、タスク分析に基づいて開発者がClaude Code WebとCLIインターフェースの選択を支援し、これらの環境間でのシームレスなセッションテレポーテーションを可能にします。Web、CLI、モバイル環境を切り替える際のセッション状態とコンテキストを管理することで、ワークフローを最適化します。様々な段階で異なるツールを必要とする複雑なプロジェクトにご活用ください。
