patent-intel
Turn patent questions into measured answers. This skill wraps a zero-dependency
CLI (scripts/patent_search.py, Python 3.8+, stdlib only) around Google Patents'
public search endpoint, so you can count, list, rank, and trend patent
publications worldwide without any API key.
The tool
python "${CLAUDE_SKILL_DIR}/scripts/patent_search.py" <command> ... [--json]
(Windows PowerShell: python "$env:CLAUDE_SKILL_DIR\scripts\patent_search.py" ...)
| Command |
What it answers |
Example |
count QUERY |
"How many patents mention X?" |
count '"agentic AI"' |
search QUERY --num 10 --sort new |
"Show me recent filings about X" |
search '"KV cache"' --sort new |
leaderboard QUERY |
"Who files the most patents about X?" |
leaderboard '"retrieval augmented generation"' |
trend QUERY --from 2019 --to 2026 |
"Is patenting in X growing?" |
trend '"large language model"' |
portfolio COMPANY |
"What does company Y patent?" |
portfolio "Anthropic" |
compare QUERY --assignees "A,B,C" |
"Who leads between A, B, C?" |
compare '"AI agent"' --assignees "OpenAI,Microsoft,Salesforce" |
report --domain SLUG --out F.md |
"Give me a landscape report on X" |
report --domain agentic-ai --out report.md |
domains |
list curated domain query packs |
|
selftest [--live] |
verify the tool works |
|
Every command takes --json for machine-readable output — prefer it when you
need to post-process results. Shared filters: --assignee, --inventor,
--country US|KR|EP|WO|..., --status GRANT|APPLICATION, --after YYYY,
--before YYYY, --date-field filing|priority|publication.
Workflow
- Map the question to a command. Vague area question ("what's happening in
AI inference patents?") → check
domains for a matching pack, then
report --domain ai-inference. Specific phrase → count/leaderboard/
search. Company question → portfolio or compare.
- Quote multi-word phrases inside the query:
'"solid state battery"',
not solid state battery (unquoted words are ANDed loosely). Full syntax:
references/search-syntax.md.
- Company names are legal entities. "Google DeepMind" finds little — use
"DeepMind Technologies". Microsoft files as "Microsoft Technology Licensing".
If a portfolio looks suspiciously small, try name variants before concluding.
- Present numbers with their meaning attached (see below), and link the
publications you cite — every search result includes a Google Patents URL.
Interpreting the data honestly
Get these right or the answer is misleading:
- Counts are patent publications, not granted patents. One invention can
appear as several publications (application + grant + family members across
countries). Say "patent publications" or "families", filter
--status GRANT when the user asks about granted patents only.
- Recent years always look smaller than they are. Applications publish
~18 months after filing, so the last ~2 years of any trend are undercounted.
Never report "filings dropped in the latest year" without this caveat.
- Leaderboard facets are sampled by the source (bucket sums can exceed the
total). Read them as relative shares — "Salesforce leads" is safe,
"Salesforce has exactly 57" is not.
- A rate-limit error is not a bug. Results are cached 24h; if the tool
reports being rate-limited, tell the user to retry in ~10-30 minutes instead
of hammering. Keep sessions under ~10 fresh (uncached) queries.
Boundaries
This is bibliographic research, not legal work. Never conclude "this doesn't
infringe", "you are free to operate", or "this idea is unpatentable" from these
results — say what the data shows (similar prior filings exist / few hits
found) and recommend a patent attorney (변리사) for legal judgments.
Going deeper
- Query syntax details: references/search-syntax.md
- Domain packs and how to extend them: references/domains.md
- Official APIs (USPTO/EPO/KIPRIS/BigQuery) for bulk or production use:
references/sources.md
1---2name: patent-intel3description: Patent research and landscape analysis over public patent data, with zero API keys. Use this skill whenever the user asks about patents or IP in any form — prior art, "has anyone patented X?", who is patenting a technology, a company's patent portfolio, patent counts, trends, landscapes, competitor IP watch, or Korean 특허 questions — even if they never say the words "patent search". Also use it when the user wants a written patent-landscape report for a technology area.4---56# patent-intel78Turn patent questions into measured answers. This skill wraps a zero-dependency9CLI (`scripts/patent_search.py`, Python 3.8+, stdlib only) around Google Patents'10public search endpoint, so you can count, list, rank, and trend patent11publications worldwide without any API key.1213## The tool1415```16python "${CLAUDE_SKILL_DIR}/scripts/patent_search.py" <command> ... [--json]17```1819(Windows PowerShell: `python "$env:CLAUDE_SKILL_DIR\scripts\patent_search.py" ...`)2021| Command | What it answers | Example |22|---|---|---|23| `count QUERY` | "How many patents mention X?" | `count '"agentic AI"'` |24| `search QUERY --num 10 --sort new` | "Show me recent filings about X" | `search '"KV cache"' --sort new` |25| `leaderboard QUERY` | "Who files the most patents about X?" | `leaderboard '"retrieval augmented generation"'` |26| `trend QUERY --from 2019 --to 2026` | "Is patenting in X growing?" | `trend '"large language model"'` |27| `portfolio COMPANY` | "What does company Y patent?" | `portfolio "Anthropic"` |28| `compare QUERY --assignees "A,B,C"` | "Who leads between A, B, C?" | `compare '"AI agent"' --assignees "OpenAI,Microsoft,Salesforce"` |29| `report --domain SLUG --out F.md` | "Give me a landscape report on X" | `report --domain agentic-ai --out report.md` |30| `domains` | list curated domain query packs | |31| `selftest [--live]` | verify the tool works | |3233Every command takes `--json` for machine-readable output — prefer it when you34need to post-process results. Shared filters: `--assignee`, `--inventor`,35`--country US|KR|EP|WO|...`, `--status GRANT|APPLICATION`, `--after YYYY`,36`--before YYYY`, `--date-field filing|priority|publication`.3738## Workflow39401. **Map the question to a command.** Vague area question ("what's happening in41 AI inference patents?") → check `domains` for a matching pack, then42 `report --domain ai-inference`. Specific phrase → `count`/`leaderboard`/43 `search`. Company question → `portfolio` or `compare`.442. **Quote multi-word phrases** inside the query: `'"solid state battery"'`,45 not `solid state battery` (unquoted words are ANDed loosely). Full syntax:46 [references/search-syntax.md](references/search-syntax.md).473. **Company names are legal entities.** "Google DeepMind" finds little — use48 "DeepMind Technologies". Microsoft files as "Microsoft Technology Licensing".49 If a portfolio looks suspiciously small, try name variants before concluding.504. **Present numbers with their meaning attached** (see below), and link the51 publications you cite — every search result includes a Google Patents URL.5253## Interpreting the data honestly5455Get these right or the answer is misleading:5657- **Counts are patent *publications*, not granted patents.** One invention can58 appear as several publications (application + grant + family members across59 countries). Say "patent publications" or "families", filter60 `--status GRANT` when the user asks about granted patents only.61- **Recent years always look smaller than they are.** Applications publish62 ~18 months after filing, so the last ~2 years of any trend are undercounted.63 Never report "filings dropped in the latest year" without this caveat.64- **Leaderboard facets are sampled** by the source (bucket sums can exceed the65 total). Read them as relative shares — "Salesforce leads" is safe,66 "Salesforce has exactly 57" is not.67- **A rate-limit error is not a bug.** Results are cached 24h; if the tool68 reports being rate-limited, tell the user to retry in ~10-30 minutes instead69 of hammering. Keep sessions under ~10 fresh (uncached) queries.7071## Boundaries7273This is bibliographic research, not legal work. Never conclude "this doesn't74infringe", "you are free to operate", or "this idea is unpatentable" from these75results — say what the data shows (similar prior filings exist / few hits76found) and recommend a patent attorney (변리사) for legal judgments.7778## Going deeper7980- Query syntax details: [references/search-syntax.md](references/search-syntax.md)81- Domain packs and how to extend them: [references/domains.md](references/domains.md)82- Official APIs (USPTO/EPO/KIPRIS/BigQuery) for bulk or production use:83 [references/sources.md](references/sources.md)
kevin9327