Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Prepare evidence-backed patent invention disclosure materials for attorney or patent-agent review. Use when the user asks for 专利点挖掘, 技术交底书, 现有技术/查新记录, claim/patent-point support mapping, invention disclosure drafts, patent figures, or a handoff package for a software/research/project invention. This is an off-DAG engineering/IP handoff skill: it does not provide legal advice, does not submit filings, does not guarantee novelty, grant, allowance, validity or registration, and emits no Light resea
.claude/skills/light0305-light-patent-disclosure/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 489% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 396% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 57% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 73% | 0% |
Turn a real project or research result into an attorney-reviewable invention disclosure packet: what problem is solved, what technical means solve it, why it differs from nearby work, what embodiments support the breadth, what figures are needed, and what counsel still needs to decide.
Read references/patent-resource-map.md before jurisdiction- or filing-sensitive work. It records the peer skills, official sources, borrowed mechanisms and honest boundaries. Read references/patent-interview-and-search.md before patent-point mining, public search, claim-ladder drafting or attorney handoff. It contains the detailed interview, search and QC rules.
READY_TO_FILE.Deliver only DRAFT, NEEDS_USER_INPUT, or READY_FOR_ATTORNEY_REVIEW.
ownership, freedom to operate, or registration outcome.
disclosure, foreign-filing rule, or priority fact is not verified, write UNKNOWN, PLANNED, or UNAVAILABLE with the next check.
papers, experiment logs, issue discussions or user-supplied records. Keep relative locators and SHA-256. Do not invent implementation details.
Graphviz, PlantUML or SVG. Do not use AI-generated bitmap images for patent drawings.
STAGE_GATES, noROUTES, no light.findings.v1, and no scientific back-edge.
Ask for or mark UNKNOWN:
US nonprovisional, PCT, EP, or undecided;
posters, standards submissions, and their dates;
Record risk_triage for public disclosure, ownership/inventorship, foreign-filing/secrecy and trade-secret redaction. Stop for the user when a public disclosure, ownership dispute, foreign filing strategy, secrecy review, or filing deadline could change the next action.
Scan only files placed in scope. For each source artifact record:
id, relative path, sha256, freshness/date if known;sharing with outside counsel.
If the source is a paper, product demo, notebook, API contract, dataset or diagram, bind the exact locator. Do not let chat memory become evidence.
For each candidate patent point, write:
control loop, signal processing, model pipeline, hardware arrangement or UI interaction rule;
Prefer one strong, defensible invention story over a pile of vague features. Ask targeted questions for tacit knowledge that the repo cannot show. Keep a short inventor interview log: problem, failed alternatives, key insight, constraints, contributors, disclosure dates and known prior art.
Search the relevant official or public sources available in the current environment, then record:
VERIFIED, PLANNED, UNKNOWN, or UNAVAILABLE.Keep inventor_known_prior_art separate from prior_art. The former is what the team already knows; the latter is the agent-run public search log. A verified public search records searched_sources; if non-patent literature is not searched, write why.
Do not call this a legal novelty opinion. If search coverage is shallow, say so and list the missing source or professional search still needed.
Before drafting final sections, write a claim strategy:
cases and alternatives;
If the broad point is unsupported, narrow it or mark it as counsel question.
Use a plain, editable structure:
Claims are only drafts for review. Keep terminology consistent with the description and avoid over-broad elements unsupported by artifacts.
Create a packet following templates/patent-disclosure-packet.example.json, then run:
bashpython scripts/disclosure_gate.py --packet patent-disclosure-packet.json --base <project-root> --as-of 2026-07-05 python scripts/disclosure_gate.py --selftest
The gate must pass before saying the packet is ready for attorney review. It checks risk triage, inventor-known prior art, public search coverage, claim ladder, support mapping, figure source, QC flags and anti-overclaim language. A failed gate means fix the packet, lower the claim, or mark the missing fact.
ACT:
sources;
ASK:
NEVER:
behalf;
alone;
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 40,988 | 6,619 | -84% | 1 | 1 | 0% | 6,240 | 2,050 | -67% | 0 | 0 | — |
case-02 | fail→fail | 37,782 | 37,232 | -1% | 1 | 1 | 0% | 6,244 | 7,979 | +28% | 0 | 0 | — |
case-03 | fail→fail | 42,603 | 7,862 | -82% | 1 | 1 | 0% | 6,235 | 2,183 | -65% | 0 | 0 | — |
case-04 | fail→pass | 3,186 | 4,940 | +55% | 1 | 1 | 0% | 414 | 2,437 | +489% | 0 | 0 | — |
case-05 | pass→pass | 14,615 | 6,300 | -57% | 1 | 1 | 0% | 2,030 | 2,604 | +28% | 0 | 0 | — |
case-06 | fail→pass | 8,043 | 23,183 | +188% | 1 | 1 | 0% | 1,086 | 5,382 | +396% | 0 | 0 | — |
case-07 | pass→pass | 13,678 | 14,911 | +9% | 1 | 1 | 0% | 2,157 | 4,015 | +86% | 0 | 0 | — |
case-08 | fail→fail | 13,871 | 14,756 | +6% | 1 | 1 | 0% | 1,965 | 4,070 | +107% | 0 | 0 | — |
case-09 | pass→pass | 11,020 | 13,595 | +23% | 1 | 1 | 0% | 1,456 | 3,517 | +142% | 0 | 0 | — |
case-10 | fail→pass | 9,661 | 3,547 | -63% | 1 | 1 | 0% | 1,452 | 2,279 | +57% | 0 | 0 | — |
case-11 | pass→pass | 13,366 | 15,441 | +16% | 1 | 1 | 0% | 2,011 | 4,035 | +101% | 0 | 0 | — |
case-12 | fail→pass | 20,713 | 17,431 | -16% | 1 | 1 | 0% | 3,448 | 4,489 | +30% | 0 | 0 | — |
case-13 | pass→pass | 14,407 | 11,636 | -19% | 1 | 1 | 0% | 2,188 | 3,541 | +62% | 0 | 0 | — |
case-14 | fail→pass | 13,983 | 11,227 | -20% | 1 | 1 | 0% | 1,967 | 3,396 | +73% | 0 | 0 | — |
case-15 | fail→pass | 10,045 | 6,223 | -38% | 1 | 1 | 0% | 1,255 | 2,701 | +115% | 0 | 0 | — |
case-16 | pass→pass | 12,201 | 9,367 | -23% | 1 | 1 | 0% | 1,852 | 3,178 | +72% | 0 | 0 | — |
case-17 | fail→pass | 6,336 | 1,622 | -74% | 1 | 1 | 0% | 951 | 1,964 | +107% | 0 | 0 | — |
case-18 | pass→pass | 14,575 | 11,719 | -20% | 1 | 1 | 0% | 2,269 | 3,514 | +55% | 0 | 0 | — |
case-19 | fail→pass | 10,544 | 8,599 | -18% | 1 | 1 | 0% | 1,572 | 3,167 | +101% | 0 | 0 | — |
case-20 | pass→pass | 6,694 | 4,821 | -28% | 1 | 1 | 0% | 945 | 2,481 | +163% | 0 | 0 | — |
case-21 | pass→pass | 10,814 | 6,110 | -43% | 1 | 1 | 0% | 1,581 | 2,637 | +67% | 0 | 0 | — |
case-22 | fail→pass | 20,612 | 18,218 | -12% | 1 | 1 | 0% | 3,478 | 4,323 | +24% | 0 | 0 | — |
case-23 | fail→pass | 18,371 | 6,212 | -66% | 1 | 1 | 0% | 3,237 | 2,678 | -17% | 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. 23 cases were attempted, and 21 counted toward the lift figure. The other 2 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 +43 percentage points is the difference between those two pass rates over the 21 comparable cases. 1 case got worse with the skill loaded, and it is 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.