recipe-prior-art-search
Skill1 starsChanged 4 months agoMITView on GitHub
Comprehensive prior-art / novelty search before filing or on a published patent — per-element synonym rows, dual EPO/USPTO patent search with a classification-only track and a citation track, an NPL sweep (OpenAlex first), a recall check against the Examiner Baseline, and X/Y/A tagging of the closest hits. Trigger when the user asks to find prior art for an invention, run a novelty search, or check what already exists before filing.
In this repository34 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- flowleap-academicSkill
- flowleap-authSkill
- flowleap-citationSkill
- flowleap-keysSkill
- flowleap-legalSkill
- flowleap-nplSkill
- flowleap-opsSkill
- flowleap-patentSkill
- flowleap-patstat-graphSkill
- flowleap-patstatSkill
- flowleap-sharedSkill
- flowleapSkill
- flowleap-toolsSkill
- flowleap-usptoSkill
- persona-ip-analystSkill
- persona-patent-attorneySkill
- persona-researcherSkill
- persona-startup-founderSkill
- recipe-academic-literature-reviewSkill
- recipe-audit-reportSkill
- recipe-claim-analysisSkill
- recipe-claim-draftingSkill
- recipe-custom-dashboardSkill
- recipe-find-betterSkill
- recipe-freedom-to-operateSkill
- recipe-infringement-chartingSkill
- recipe-invalidity-analysisSkill
- recipe-invention-disclosureSkill
- recipe-maintenance-feesSkill
- recipe-office-action-responseSkill
- recipe-patent-landscapeSkill
- recipe-patent-to-reportSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 5 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 40
When the subject is a published document, read its full cpc[] from flowleap --json ops biblio <publication>, including the 2000-series
- Very specific: line 48
Verify any CPC code against the official scheme (flowleap patstat query on flowleap.cpc_scheme) — never guess codes.
- Distinctive: line 72
Run both databases — they are not interchangeable:
- Distinctive: line 170
Read past the abstract (the open-access URL or DOI) before relying on a paper.
- Distinctive: line 211
When the subject is a published patent or application, or the disclosure names one, check the search against what the examiners cited, …
- Distinctive: line 264
Never judge a result set from its first page: read it in full or narrow it until you can.
- Distinctive: line 266
Always run the Step 5 recall check when the subject has an Examiner Baseline, and report "found n of m".
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Recipe: Prior Art Search
A multi-step workflow for comprehensive prior-art searching. Each database uses
its own query syntax — see flowleap-uspto for the USPTO Lucene grammar.
In a chat client with the FlowLeap connector or in the FlowLeap Patent AI app, call the tools named in the flowleap-shared tool table instead of these commands, and never run a flowleap command there. The steps stay the same.
Steps
Step 1: Write the Search Queries
You write both queries yourself — the invention description never leaves the
machine to become a query. The method (from flowleap-patent, where the CQL
fields live):
- Extract the terms. List every specific noun phrase in the invention
description; for each one you leave out of the query, state why. Unsure of
a phrasing ("glass ceramic" vs "glass-ceramic")? OR both forms — never drop
the term.
Build one synonym row per claim element, not per invention: the
claim's own word, the structural equivalents another drafter would use,
and the functional words for what the element does. Older art and other
offices describe the same part differently, and a document that uses none
of your words is invisible to every keyword query. Example (hydraulic
actuator, EP4758353A1): pump: bidirectional, reversible, tiltable,
over-centre, four-quadrant; chamber: cap side, rod side, piston side,
annulus; pressure reservoir: accumulator, tank, pressure compensator. OR
within one row (
(ta=a OR ta=b)); never let one element's word stand in for the whole invention.ta=matches the English title and abstract, so English synonyms come first; add German or French terms when the scope includes DE or FR documents that may carry only original-language text. When the subject is a published document, read its fullcpc[]fromflowleap --json ops biblio <publication>, including the 2000-series indexing codes (e.g.F15B2211/7053): they describe the features independent of wording, and they feed the classification-only track in Step 2. - Write the CQL around a discriminating term — the specific subject
matter, never just the technology area, never a CPC class alone. Grouping
repeats the field:
(ta=X OR ta=Y)is valid;ta=(X OR Y)is a hard OPS 404. Verify any CPC code against the official scheme (flowleap patstat queryonflowleap.cpc_scheme) — never guess codes. - Probe the count — mandatory.
patent searchshows no total, so run the same tool directly and readtotalfrom its payload.details=falseskips the per-document bibliography fan-out, so the probe stays cheap:
flowleap --json tools run search_patents query='<self-written CQL>' range=1-1 details=falseOver ~1,000 hits: add the next discriminating term from your extraction list. Under 10: broaden — this is a novelty pass, so start broad (drop the classification, OR in synonyms) and narrow from the count, never the reverse.
The USPTO leg uses ODP Lucene over title + metadata (see flowleap-uspto):
same term extraction, same probe discipline, but the discrimination must be a
term that plausibly appears in an invention title.
Show the statement as a Search Statement block — see flowleap-patent references/search-statement.md.
Done when you have one probed EPO CQL query and one USPTO ODP query.
Step 2: Search Patents (EPO + USPTO)
Run both databases — they are not interchangeable:
patent searchhits EPO OPS: the worldwide bibliographic tier (DOCDB), including EPO's own bibliographic copy of US documents.--countries USnarrows this EPO collection to US members; it does not reach USPTO's Open Data Portal.uspto searchhits USPTO ODP directly, returning US application and prosecution metadata (thepatentFileWrapperDataBag) that the EPO copy lacks.
For US prior art the USPTO leg must go through USPTO ODP; do not substitute
patent search --countries US (that returns EPO's copy, not ODP records).
EPO: the CQL from step 1 is a query string:
flowleap --json patent search --query "<CQL from step 1>" --limit 20USPTO: the Lucene query from step 1 goes in --query (or wrap it in a
full ODP request body via --body when you need fields/enrich):
flowleap --json uspto search --query '<ODP Lucene from step 1>' --limit 20USPTO recall caveat — ODP is title-only. ODP search indexes the invention title and a few metadata fields; there is no abstract or claims full-text. A distinguishing feature that only appears in the abstract (e.g. "UV-C sterilization") therefore cannot be matched, and a query that ANDs such a term onto the search returns 0. When the USPTO leg comes back empty (the CLI prints a note and auto-retries once without the CPC filter), run a title recall pass on the core device noun and triage abstracts afterwards:
# Recall on the device category, with singular/plural variants; drop the
# abstract-only qualifier. Then read abstracts to keep the on-point hits.
flowleap --json uspto search --query 'applicationMetaData.inventionTitle:earbuds AND applicationMetaData.inventionTitle:"charging case"' --limit 25
flowleap ops abstract <application-or-publication-number>If one office answers provider_keys_required, its patent-data key is
missing. Do not narrow the search to the office whose key happens to be set, and
do not fill the gap with web-scraped results: run the live office in full,
report the other as a missing-key gap (not "no prior art found" and not "limited
coverage"), and ask for the free key at the end. See flowleap-keys.
Run at least two independent tracks, so a document one misses another can catch:
- Keyword track: the queries above, built from the per-element rows.
- Classification-only track: the main group AND combinations of 2-3
indexing codes that map to essential elements, with the date limit and no
keywords, e.g.
cpc=F15B7/006 AND cpc=F15B2211/7053 AND cpc=F15B2211/785 AND pd<20230810(45 hits, probed live 2026-10-11). A single code is a neighbourhood; a conjunction of specific indexing codes is a set small enough to read in full, and it finds documents whose wording no keyword query matched. Try several combinations; log each with its count. - Citation track (Step 4): one hop from the closest hits.
Read the whole set, never page 1. Results come back newest first, not by
relevance: older art sits at the end. patent search --limit stops at 100, so
page a larger set with ops search (--start 101 --end 200, …). When a set is
too large to read in full (over about 300), narrow it (the next synonym row,
another indexing code, a date window) until it is, and log each step. Never
judge a set, or record a document as absent, from its first page.
OPS returns one publication per family: a hit may be the CA, WO or JP
member of the US document you know (US2013333894A1 comes back as
CA2828987A1). Compare documents by family (ops family), not by number.
flowleap --json ops search --cql '<classification-only CQL>' --start 101 --end 200Done when both databases have returned ranked results, or one is reported as an open missing-key gap, and each set you rely on has been read in full.
Step 3: Search Non-Patent Literature
flowleap --json npl "<discriminating terms>" --to-year <critical-year> --limit 25
flowleap --json academic search "<discriminating terms>" --to-year <critical-year> --limit 15npl(OpenAlex, 250M+ works) first. Reuse the discriminating terms from the per-element rows in paper wording, run a variant per essential feature, and check eachpublicationDateagainst the exact critical date. Narrow with--type(proceedings-article,dissertation,standard, …) when a track calls for it. Log each query and its hit count.academic search(Semantic Scholar + arXiv) as a second pass for preprints and CS/physics work; dedupe against step 1 by DOI or title. A source named insources[]asrate_limitedorerroris a gap, not "no literature".- Papers an office already cited (the
nplentries of a source or family publication'scitedReferences) are resolved, not searched:flowleap --json ops biblio <publication> --resolve-nplresolves all of them,flowleap --json npl-work "<cited string or DOI>"one. Only amatchedrecord is the cited paper. - Read past the abstract (the open-access URL or DOI) before relying on a paper; cite it by authors, title, venue, date and DOI.
Step 4: Deep Dive on the Closest Hits
Deep-dive every hit whose abstract maps to at least one independent feature of the invention — at minimum the top 5 by rank:
flowleap ops abstract <patent-number>
flowleap ops claims <patent-number>
flowleap ops family <patent-number>For an office with no text source, use the family-member workaround in flowleap-shared.
One hop from the closest hits (the citation track): for each closest
candidate of the keyword and classification tracks, and for each Examiner
Baseline document once Step 5 has it, read its cited references (ops biblio
citedReferences[]) and its forward citations (citation forward). Keep the
documents published before the cutoff that share a CPC main group or indexing
code with the subject. When a disclosure names a source or family
publication, work its cited references first (for an EP family they sit on the
A3 search-report publication); those documents are of record, not found by
your search: keep them apart for the recall check.
flowleap --json ops biblio <closest-hit> # citedReferences[]
flowleap --json citation forward <closest-hit> --size 50Record how deeply each document was read: bibliography/abstract only or
full text read (claims and description). Map an element to a passage only
after a full-text read; an abstract-only document can be named as a
candidate, never as supporting a row.
Done when each qualifying hit has its claims and family pulled and the citation hop has run.
Step 5: Recall Check against the Examiner Baseline
When the subject is a published patent or application, or the disclosure names one, check the search against what the examiners cited, after the keyword, classification and citation tracks have run:
flowleap --json patent examiner-baseline <publication> # any family member- Take every document an examiner cited (X, Y and A categories, ISR and national searches); applicant-cited documents are listed separately.
- A Baseline document counts as found only when a search query returned
it or a member of its family. A
pn=lookup, or reading the source's own cited references, does not count. - For each miss, find out why: compare its CPC codes and abstract wording with your queries (a synonym or indexing code the queries lacked, or a page you did not read). Run the query that fixes the cause and log it with its purpose. Repeat until each miss is found by a search or explained.
- State the result as "found n of m examiner-cited documents by search", list each miss with its reason, and put unexplained misses under limitations. A low recall means the queries are too narrow for new art too: say so rather than presenting new candidates as complete.
Then look for art beyond the Baseline, starting from the claim elements the examiner documents do not cover.
Step 6: Tag Each Reference X / Y / A
Against the invention's features, tag every retained reference:
- X — alone anticipates a feature (novelty-destroying)
- Y — anticipates only in combination with another reference
- A — general background
Done when every retained reference carries an X/Y/A tag, X-tagged first.
Step 7: Map the X References Element by Element
For each X-tagged reference, show the anticipation per element rather than asserting it per document. One table, one row per invention feature (or claim element, when the user supplied a claim), one column per X reference:
| Claim element | <Reference 1> | <Reference 2> |Each cell quotes the disclosing passage from the retrieved claims — original language plus a translation where applicable. A feature no X reference discloses is the novelty candidate; say so under the table.
Done when every element row is either quoted from at least one X reference or named as a potential point of novelty.
Rules
- Never judge a result set from its first page: read it in full or narrow it until you can.
- Always run the Step 5 recall check when the subject has an Examiner Baseline, and report "found n of m".
Output
A prior-art table with:
- One row per patent family (the closest member represents the family; use
ops familyto collapse duplicates) - Each row tagged X / Y / A, X-tagged references surfaced first
- The element-by-element mapping table for the X references (Step 7)
- Papers on the same topic, cited by authors, title, venue, date and DOI
- The tracks run, each query with its count, and the review depth per document
- The recall result "found n of m examiner-cited documents by search", with each miss and its reason, when the subject has an Examiner Baseline
- Claims and abstracts from the closest prior art
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

