Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Search patent databases and academic literature for prior art relevant to an invention. Use when user says "现有技术检索", "prior art search", "专利检索", "check patents", or wants to find relevant prior art.
.claude/skills/wanshuiyin-prior-art-search/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-22 | ✗→✓ | ▲ Improved | 54% | 0% |
| case-04 | ✓→✗ | ▼ Worse | 16% | 0% |
| case-05 | ✓→✗ | ▼ Worse | -36% | 0% |
| case-06 | ✓→✗ | ▼ Worse | -32% | 0% |
| case-07 | ✓→✗ | ▼ Worse | -45% | 0% |
Search patents and literature for prior art relevant to: $ARGUMENTS
Adapted from /research-lit for patent-specific searching.
MAX_PATENT_RESULTS = 20 — Maximum patent documents to analyze in detailMAX_PAPER_RESULTS = 15 — Maximum academic papers to analyze in detailSEARCH_YEARS = 10 — How many years back to searchPATENT_DATABASES = "google-patents, espacenet" — Patent databases to searchRead the invention description from:
$ARGUMENTS if it contains technical detailspatent/INVENTION_BRIEF.md if it existsINVENTION_BRIEF.md if it exists at project rootLoad ../shared-references/prior-art-databases.md for search strategy templates and IPC/CPC classification guidance.
From the invention description, identify:
For EACH search concept, search via:
Google Patents (via WebSearch):
WebSearch: "site:patents.google.com [keywords]"
WebSearch: "[keywords] patent"Espacenet (via WebFetch):
Assignee/Inventor Search:
For each potentially relevant patent found:
Search the same concepts in academic databases:
WebSearch "[keywords] site:scholar.google.com"/arxiv if available, or WebSearch): Search for preprints/semantic-scholar if API key set, or WebSearch)For each relevant paper found:
For each reference found, assess:
Organize results by IPC/CPC classification to see the technical landscape.
Based on the search results:
Disclaimer: This is a preliminary assessment only. A professional freedom-to-operate analysis by a patent attorney is recommended before filing.
Write patent/PRIOR_ART_REPORT.md with:
markdown## Prior Art Search Report ### Invention Summary [1-2 sentence description of the searched invention] ### Search Strategy - Keywords used: [...] - IPC/CPC classes searched: [...] - Databases searched: Google Patents, Espacenet, Google Scholar, arXiv - Date range: [year] to present ### Patent References Found | # | Patent No. | Title | Date | Assignee | IPC/CPC | Key Teaching | Overlap Risk | |---|-----------|-------|------|----------|---------|-------------|-------------| | 1 | CN... / US... | [title] | [date] | [assignee] | [codes] | [2-3 sentences] | HIGH/MEDIUM/LOW | ### Non-Patent Literature Found | # | Reference | Title | Authors/Venue | Year | Key Contribution | Relevance | |---|-----------|-------|--------------|------|-----------------|-----------| | 1 | [DOI/link] | [title] | [authors] | [year] | [1-2 sentences] | HIGH/MEDIUM/LOW | ### Prior Art Landscape [Organized by technical approach or IPC class, not just chronological] ### Freedom-to-Operate Preliminary Assessment [Which existing patents might block the invention? What is the risk level?] ### Recommendations - Suggested claim scope adjustments based on prior art - Areas where novelty appears strongest - References to watch during prosecution
[VERIFY].| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 35,065 | 8,274 | -76% | 1 | 1 | 0% | 6,238 | 1,471 | -76% | 0 | 0 | — |
case-02 | fail→fail | 28,070 | 7,281 | -74% | 1 | 1 | 0% | 4,933 | 1,614 | -67% | 0 | 0 | — |
case-03 | fail→fail | 48,329 | 10,261 | -79% | 1 | 1 | 0% | 5,377 | 1,422 | -74% | 0 | 0 | — |
case-04 | pass→fail | 7,296 | 8,189 | +12% | 1 | 1 | 0% | 1,560 | 1,803 | +16% | 0 | 0 | — |
case-05 | pass→fail | 19,567 | 50,404 | +158% | 1 | 1 | 0% | 3,357 | 2,165 | -36% | 0 | 0 | — |
case-06 | pass→fail | 15,167 | 23,074 | +52% | 1 | 1 | 0% | 2,560 | 1,737 | -32% | 0 | 0 | — |
case-07 | pass→fail | 19,158 | 35,209 | +84% | 1 | 1 | 0% | 3,441 | 1,903 | -45% | 0 | 0 | — |
case-08 | fail→fail | 21,805 | 7,494 | -66% | 1 | 1 | 0% | 3,557 | 1,477 | -58% | 0 | 0 | — |
case-09 | fail→fail | 24,120 | 8,521 | -65% | 1 | 1 | 0% | 4,152 | 1,526 | -63% | 0 | 0 | — |
case-10 | pass→fail | 26,648 | 4,866 | -82% | 1 | 1 | 0% | 4,109 | 1,450 | -65% | 0 | 0 | — |
case-11 | pass→fail | 31,771 | 8,000 | -75% | 1 | 1 | 0% | 6,199 | 1,612 | -74% | 0 | 0 | — |
case-12 | pass→fail | 31,699 | 4,977 | -84% | 1 | 1 | 0% | 6,202 | 1,639 | -74% | 0 | 0 | — |
case-13 | fail→fail | 3,226 | 6,371 | +97% | 1 | 1 | 0% | 262 | 1,771 | +576% | 0 | 0 | — |
case-14 | fail→fail | 3,569 | 4,014 | +12% | 1 | 1 | 0% | 198 | 1,415 | +615% | 0 | 0 | — |
case-15 | fail→fail | 26,256 | 8,941 | -66% | 1 | 1 | 0% | 4,281 | 1,881 | -56% | 0 | 0 | — |
case-16 | fail→fail | 26,893 | 32,526 | +21% | 1 | 1 | 0% | 4,528 | 2,900 | -36% | 0 | 0 | — |
case-17 | pass→fail | 26,044 | 7,073 | -73% | 1 | 1 | 0% | 3,664 | 2,058 | -44% | 0 | 0 | — |
case-18 | pass→fail | 23,944 | 10,633 | -56% | 1 | 1 | 0% | 4,186 | 2,178 | -48% | 0 | 0 | — |
case-19 | pass→fail | 33,205 | 2,436 | -93% | 1 | 1 | 0% | 6,187 | 1,448 | -77% | 0 | 0 | — |
case-20 | fail→fail | 30,811 | 7,441 | -76% | 1 | 1 | 0% | 5,785 | 1,632 | -72% | 0 | 0 | — |
case-21 | pass→fail | 6,237 | 5,983 | -4% | 1 | 1 | 0% | 1,157 | 1,545 | +34% | 0 | 0 | — |
case-22 | fail→pass | 23,520 | 30,767 | +31% | 1 | 1 | 0% | 3,872 | 5,974 | +54% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 1 counted toward the lift figure. The other 21 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of -45 percentage points is the difference between those two pass rates over the 1 comparable cases. 14 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.