nomhiro/patent-investigator

Patent Investigator Skill

Conduct a structured prior art search using only free services, then deliver a patentability assessment report with a claim chart.

Prerequisites

The only thing that needs to be set up in advance is an EPO OPS API key (free).

| Item | How to get it | |---|---| | EPO OPS key | Register at https://developers.epo.org (free, "Non-paying" plan, ~5 min). Set env vars EPO_KEY and EPO_SECRET. |

If the env vars are missing when you run the helper script, it will print an error with the registration URL. In that case, tell the user what they need to do and stop gracefully.

Tools you will use

| Tool | Purpose | |---|---| | web_search | Google Patents / general web patent search (JP + EN) | | web_fetch | Retrieve full patent pages from Google Patents / Espacenet | | bash_tool | Run scripts/epo_ops_search.py for EPO API + Semantic Scholar | | Your own reasoning | Claim decomposition, comparison, novelty assessment |

The helper script lives at scripts/epo_ops_search.py inside this skill folder. Read it once at the start to confirm the path, then call it via bash.

Helper script commands

python <skill>/scripts/epo_ops_search.py search   "CQL query"
python <skill>/scripts/epo_ops_search.py biblio    <CC> <number> <kind>
python <skill>/scripts/epo_ops_search.py claims    <CC> <number> <kind>
python <skill>/scripts/epo_ops_search.py abstract  <CC> <number> <kind>
python <skill>/scripts/epo_ops_search.py description <CC> <number> <kind>
python <skill>/scripts/epo_ops_search.py family    <CC> <number> <kind>
python <skill>/scripts/epo_ops_search.py citations <CC> <number> <kind>
python <skill>/scripts/epo_ops_search.py scholar   "query" [limit]

CQL examples (EPO OPS search syntax):

• ti=interview AND ta=real-time AND ic=G06F
• pa=Toyota AND ti=support system
• ti="mind map" AND ta="speech recognition"

Workflow

Follow these six phases in order. Do NOT skip Phase 1's confirmation step.

---

Phase 1 — Structure the invention

Turn the user's description into a structured invention definition:

1. Draft independent claim — write it in patent claim style: "〜であって、Aと、Bと、Cとを備える、〜システム。" (or English equivalent)
2. Decompose into elements — label them A, B, C, …
3. Generate search keywords
   • 3–5 Japanese keyword sets (with synonyms / broader terms)
   • 3–5 English keyword sets
4. Estimate classification codes — suggest likely IPC/CPC codes (e.g. G06F, H04L)
5. List competitor candidates — companies likely filing in this space

Present all of the above to the user and ask:

「この構成要件と検索方針で調査を進めてよいですか？修正があれば教えてください。」

Wait for confirmation before moving on. This checkpoint is critical because a wrong decomposition will send all downstream searches in the wrong direction.

---

Phase 2 — Multi-axis search

Execute at least 6 searches across the following channels. More is better — vary keywords and axes to maximise recall.

A. Google Patents (web_search — 2–3 queries)

• English keywords: "real-time transcription" "mind map" AI advice patent
• Japanese keywords: リアルタイム 文字起こし マインドマップ AI 特許
• Element combinations: search for A+B, A+C, B+C separately

B. EPO OPS structured search (bash_tool — 2–3 queries)

• CPC × keyword: python …/epo_ops_search.py search "ti=… AND ta=… AND ic=…"
• Applicant search: python …/epo_ops_search.py search "pa=CompanyName AND ti=…"

C. Scholarly literature (bash_tool — 1 query)

• python …/epo_ops_search.py scholar "real-time interview AI support" 10
• In AI/software domains, papers on arXiv often constitute prior art.

D. PQAI semantic search (web_search — 1 query, optional)

• Search for site:projectpq.ai <English invention summary> or use web_fetch on https://search.projectpq.ai/ (results limited to 20 but semantically ranked — useful as a cross-check).

Log every query you run (search string + source) so you can report them later.

---

Phase 3 — First-pass screening

1. Deduplicate results by patent number / family ID.
2. For each result, read the title + abstract.
3. Rate relevance to the invention: High / Medium / Low.
4. Keep the top 10–15 High/Medium candidates for detailed analysis.
5. Briefly tell the user how many hits you got and how many survived screening.

---

Phase 4 — Claim-level analysis (this is where quality is made or lost)

For each High-relevance candidate:

1. Get the full independent claims:
   • First try web_fetch on the Google Patents page (https://patents.google.com/patent/<patent_number>)
   • If that fails or is incomplete, use the helper script: python …/epo_ops_search.py claims <CC> <number> <kind>
2. Extract the claim elements of the prior art's independent claim(s).
3. Map each element against the invention's elements A, B, C, …
   • ○ = disclosed
   • △ = partially disclosed or different scope
   • × = not disclosed
4. Assign a threat level:
   • A (Critical): Every element is ○ — this single reference may destroy novelty.
   • B (Watch): Most elements ○ or △ — could be combined with another reference to challenge inventive step.
   • C (Low risk): One or more key elements are × — unlikely to be a blocking reference on its own.

---

Phase 5 — Citation chain exploration (conditional loop)

If any threat-A or threat-B references were found in Phase 4:

1. Retrieve their citation lists: python …/epo_ops_search.py citations <CC> <number> <kind>
2. Screen the cited / citing patents (Phase 3 logic).
3. Run Phase 4 on any new High-relevance finds.
4. Repeat up to 2 times (avoid runaway loops).

This mimics what a professional searcher does: follow the citation graph to uncover related art that keyword search alone would miss.

---

Phase 6 — Report

Produce the final report with these sections. Use the user's language (Japanese if the conversation is in Japanese, English otherwise).

6-1. 調査概要 / Search summary

• Invention summary (the draft independent claim from Phase 1)
• Databases searched, queries used, date of search
• Hit count → screened count → detailed analysis count

6-2. 構成要件対比表 / Claim chart

| 構成要件 | 先行技術 1 | 先行技術 2 | 先行技術 3 | |---|---|---|---| | A: … | ○ … | △ … | × | | B: … | ○ … | ○ … | ○ … | | C: … | × | × | △ … |

Include a one-line note for each cell explaining why it's ○/△/×.

6-3. 主要先行技術の詳細 / Key references

For each threat-A/B reference:

• Patent number, title, applicant, filing date, link
• Technical summary (2–3 sentences)
• Similarities and differences vs. the invention
• Threat level + rationale

6-4. 特許性の予備評価 / Preliminary patentability assessment

• Novelty (新規性): Is there a single reference disclosing all elements?
• Inventive step (進歩性): Could a skilled person combine 2–3 references to arrive at the invention?
• Conclusion: one paragraph with the overall assessment.

6-5. 出願に向けた提案 / Filing recommendations

• Which elements to emphasise in claims
• Possible dependent claims to strengthen the position
• Areas where additional search may be warranted
• Reminder that this is a preliminary AI assessment — professional review by a patent attorney (弁理士) is recommended before filing.

---

Important caveats to always include

• This search is preliminary and not exhaustive.
• Chinese and Korean patent coverage is limited in the free databases used.
• Japanese patent full-text claims may only be available via machine translation (Google Patents).
• The assessment is generated by an AI and does not constitute legal advice.
• Always recommend consulting a patent attorney (弁理士) for filing decisions.