patent_prior_art v1.0.0

name: "patent_prior_art" description: "Run prior-art and patent-landscape research on primary databases: novelty searches on Google Patents and Espacenet with documented queries, CPC classification follow-up, INPADOC family resolution, citation chaining, and preliminary freedom-to-operate flags. Triggers when an agent needs to check whether an idea is novel, map who owns a technology area, or flag infringement risk before building." metadata: { "includeInPrompt": true }

Patent Prior Art

Purpose

Teach an agent to do real prior-art research the way a patent professional starts: structured queries on primary databases (Google Patents, Espacenet), classification-driven expansion via CPC, family and citation follow-up, and a written search log that makes every finding reproducible. Two different questions, two different searches: novelty (can this be patented?) and freedom to operate (can this be sold?). This skill is preliminary research guidance, not legal advice. A real FTO opinion or invalidity conclusion needs a qualified patent attorney.

Workflow

1. Name the question first. Novelty asks whether an invention is new before filing, so any disclosure anywhere before the filing date counts and legal status does not matter. Freedom to operate asks whether a product can be commercialized without infringing, so only active rights count (granted patents with fees paid, pending applications that could grant) and only in the jurisdictions where the product will be made or sold. Run the right search for the question. Passing one says nothing about the other.

1. Decompose the invention. Break it into its functional concepts: what problem it solves, how it solves it, the distinguishing feature. List 3 to 5 synonyms and alternative technical terms for each concept, including older terminology. Old patents describe mechanisms with different language, so one vocabulary is never enough.

1. Keyword search on Google Patents. Use the Advanced Search page for Boolean logic. Build each concept as an OR group of synonyms, then AND the groups: (vector OR embedding) AND (index OR database) AND (cache OR memory) AND (retrieve OR search). Rules that matter: the default operator is AND with left associativity, so safety OR seat belt reads as (safety OR seat) AND belt; always parenthesize. Quotes force exact phrases. Each word gets plurals and synonyms automatically. Full operator table lives in references/query-cheatsheet.md. For novelty, add before:priority:YYYYMMDD so results predate the filing date. Scope claims with CL=(...) when you need precision: a hit in the claims is worth ten in the description.

1. Harvest CPC codes from the best hits. Take the 5 to 10 most relevant results, read their CPC classifications, and find the codes that recur. Then search the code directly, alone first, then refined with keywords. CPC is language-independent and examiner-assigned, so it catches patents your keywords never would. Use the Espacenet CPC browser to walk the classification tree down to the narrowest matching subclass. Details and examples in references/query-cheatsheet.md.

1. Cross-check on Espacenet. Re-run the strongest queries using Espacenet smart search field codes (ctxt= for title/abstract/claims, claims= for claims only, cpc= for classification) and proximity operators like prox/distance<5 when two concepts must appear near each other. Espacenet covers 80+ countries and is the authoritative route to INPADOC family and legal-status data. For non-US coverage this step is not optional.

1. Resolve families. For each candidate hit, open the patent family. The simple family (same priority, "also published as") gives equivalents; read one and note the rest instead of reading duplicates. The extended family (all linked via priority chains) can reveal the earliest disclosure, which may predate the document you found. Family data covers published members only and is retrieved from INPADOC via Espacenet.

1. Chain citations and check non-patent literature. Follow backward citations (what the patent built on) and forward citations (who built on it). Then search the same concepts in Google Scholar, arXiv, IEEE, or the domain's literature: papers, preprints, and conference proceedings all count as prior art. Citation chaining routinely finds the reference keyword searches miss.

1. For FTO flags: read independent claims and check status per jurisdiction. A granted patent grants the right to exclude, not the right to practise; your own patent says nothing about whether you can launch. Map each independent claim element against the product's features. Then check legal status member by member: status is country-level, so a family can be dead in one jurisdiction and alive in another. Term is 20 years from filing, and rights lapse when maintenance fees stop. Record flags as flags, not verdicts: "possible overlap, needs counsel" is the honest output of a preliminary pass.

1. Write the search log before you report. Every query you ran, the database, the date, the hit count, and which hits survived. A finding without its query is not reproducible and does not count. Use bin/log-search.py or the template in references/report-template.md.

Output Contract

• Search log: one entry per query (database, query string verbatim, date run, hit count, notes). The agent must be able to re-run every query and get the same hits.
• Hits table: publication number, title, filing/priority date, jurisdiction, relevance note, family scope (simple vs extended checked), why it matters or why it was discarded.
• Question-specific verdict: novelty search ends with the closest disclosures ranked by how much of the invention they anticipate; FTO flag pass ends with a claim-element mapping table and a per-jurisdiction status note for each live-looking patent.
• Limits stated in writing: searches run on free databases with known coverage gaps; this is preliminary research, not a legal opinion; any clearance or invalidity decision requires a patent attorney.

Operating Rules

1. Verify every factual claim the way the Accuracy Rule demands: no invented numbers, dates, quotes, or events. A legal-status claim cites the INPADOC status record it came from.
2. Never skip the question in step 1. A novelty answer must never be presented as FTO clearance, and a patent grant must never be presented as permission to build.
3. Log the query at the time you run it, not from memory afterward. Verbatim strings only.
4. Check status at the member level, never the family level. One dead member does not kill a family.
5. When a query returns thousands of hits, narrow by classification first, then keywords, then dates. When it returns nothing, suspect your vocabulary and add synonyms, do not declare the field empty.
6. Expired and abandoned patents still count as prior art. Publication before the filing date is what matters.
7. State uncertainty plainly. "No hits found on these queries" is a true sentence; "this is clear" is not, unless counsel said so.