patent-research-skill

An Agent Skill for patent prior-art search, novelty/inventiveness self-check,
evidence-chain management, and acceptance gates (G-S).
Works with Claude Code, Codex CLI, and the Codex desktop app.

This repository contains only distributable assets: the skill source, installers,
the release builder, and release notes.
No real project data, patent drafts, target IDs, or candidate IDs are included;
the sample workspace is synthetic (TGT-0002).

Skill 3.8.0 / kernel 0.9.0 — revision↔conclusions linkage (V038), 9-case benchmark, structured support citations

Three additions in this release (the P0–P4 upgrade of 3.5.0 below remains the capability base):

• V038 — revision↔conclusions linkage. A patch to the manuscript automatically marks the
  novelty conclusions stale; revise status exits 1 and revise close refuses until each
  stale conclusion is human-reconfirmed (revise reconfirm --by <name>) or explicitly waived
  into the ledger (--waive --reason). validate reports unprocessed stale conclusions as errors.
• D-019 — structured support citations. Drafting support mappings accept optional
  citations (source + L-range / [NNNN] locator) that the audit verifies against the
  workspace; unresolvable citations are errors.
• Benchmark at 9 cases (review P1-1 range 8–12): 45 features, 375 judgment points,
  140 citations; gold solutions score perfect in CI.

The P0–P4 upgrade plan of 3.5.0
(docs/patent-skill-upgrade-implementation-plan.md)
remains the capability base:

• P0 versioned data contracts + non-destructive workspace upgrade (upgrade check/apply);
• P1 the third patent-discovery skill (material intake → technical facts → patent
  candidates → human confirmation → research handoff, G-I control layer);
• P2 structured drafting (DraftingPlan → ClaimAtom → Claim → SupportMapping, G-D audit,
  concept/review_draft writing gate) and change-impact propagation;
• P3 the professional figure engine (figures plan/render/validate, deterministic SVG,
  F-001…F-014 figure-text checks);
• P4 complete delivery (frozen PatentDocumentModel → DOCX/optional PDF → delivery manifest
  with artifact SHA-256 → human sign-off bound to the manifest fingerprint → release_ready).
  CNIPA XML remains not attempted until independently researched and verified.

The kernel is reconstructed in CI from the public 0.1.2 sdist plus
release/patches/patentkit-0.1.2-to-0.9.0.patch; the full package ships all three skills.
The skill-matrix decision (2026-10-25, see
docs/skill-matrix-plan.md) still describes the split into
research and drafting (OA merged into drafting), and 3.6.0 additionally ships
patent-discovery. Everything below still describes the 2.5.0 Chinese-primary +
delivery-safety release, which 3.x inherits in full.

Skill 3.6.0 — release doctor + synthetic patent benchmark

Kernel unchanged (0.7.0 at the time). Two additions for maintainers and evaluators:

• tools/release_doctor.py — one-command release health check. Offline mode runs
  pairing points, guard scripts, a fresh dual-zip build with manifest/SUMS deep checks and
  per-platform installer smoke tests; release mode audits the actual GitHub releases
  (assets, hashes vs SHA256SUMS, released-zip manifest, notes) and reruns installer tests
  plus an isolated kernel-wheel import against the downloaded assets.
• benchmarks/ — a public benchmark of synthetic patent
  cases measuring citation accuracy (verbatim, via the kernel's own passage check),
  technical-feature omission, and misjudgment rate against per-case ground truth. Each case
  ships a participant scaffold and a reference solution; CI requires the reference
  solutions to score perfect (--gate perfect), and controlled-corruption tests ensure the
  scorer is not a rubber stamp. All publication numbers use the non-existent ZZ country
  prefix; no real application material may enter benchmarks/.

Skill 2.5.0 highlights (inherited by 3.x)

Two headline changes over 2.4.0:

1. Chinese-primary runtime. SKILL.md, all 13 references and the assets are now the
   authoritative Chinese edition (the target practice is Chinese patent work); the English
   edition lives on as references.en/ and is kept term-consistent by
   evals/check_terminology.py. A bilingual glossary (references/glossary.md) pins the
   terminology both ways.
2. Delivery-safety layer (introduced in the 2.5.0 kernel series). export-docx (workspace-verbatim .docx search
   report), map (offline self-contained HTML overview), layout (read-only 1+N portfolio
   view), ledger (evidence ledger; V034 rejects conclusions citing IDs outside it),
   provenance (numerical-claim A/B/C registry; V031/V032 fail closed), figures
   (mark consistency V033 + deterministic SVG), deadlines (statutory ledger), plus
   scripts/self_test.py for post-install self-checks.

Quality engineering added in the same release: evals/ (14 trigger/refusal/boundary/format
cases, anchored to the SKILL description), tools/check_no_workspace_data.py (secrecy guard),
secrecy .gitignore rules, and a reconstructed tests/conftest.py so the kernel's full test
suite runs from the patched tree in CI.

Everything from 2.4.0 still holds: source provenance and judgment are separate (assessment
protocol); supported AI
assessments may count for G-S-02; G-S-03 requires explicit human confirmation bound to the
current judgment. CI reconstructs the kernel from the verified public source and
patch, builds both packages, and validates installation on
Windows/Linux with Python 3.11/3.12 before publishing.

Download & install

Download the paired Skill 3.8.0 and patentkit 0.9.0 releases:

Windows (one click)

Put the zip, the wheel, and install-oneclick.cmd (extracted from the zip) in the
same folder, then double-click install-oneclick.cmd.

Manual (any platform)

python -m pip install ./patentkit-0.9.0-py3-none-any.whl
Expand-Archive ./patent-research-skill-v3.8.0.zip -DestinationPath ./pkg
pwsh -File ./pkg/install.ps1 -Client both -Force

python3 -m pip install ./patentkit-0.2.0-py3-none-any.whl
unzip patent-research-skill-v3.8.0.zip -d pkg
chmod +x pkg/install.sh && ./pkg/install.sh --client both --force

After installation, restart the client:

• Codex desktop: open Skills in the sidebar, select "Patent Research", or type @ to pick it.
• Claude Code: ask "use the patent-research skill to locate the workspace and run the self-check".

Repository layout

research/                patent-research 技能：SKILL.md + references/(zh 01-08+glossary) + references.en/ + scripts/ + assets/
drafting/                patent-drafting 技能：SKILL.md + references/(zh 00,09-17+glossary)（OA 已并入；14-17 为 P2/P3/P4 结构化撰写、附图与交付管线）
discovery/               patent-discovery 技能（P1 开发线，未发布）：SKILL.md + references/(zh 00-03+glossary) + references.en/
evals/                   trigger/refusal/boundary eval set + terminology consistency check (CI)
release/                 installers (ps1/sh/oneclick), install docs, patches, release notes
tools/                   reproducible ZIP builder + paired release publisher + secrecy guard
docs/                    design memos (skill matrix, data sources, scope decisions) + verification notes
.github/workflows/ci.yml  wheel/ZIP builds, regressions, evals, guards, installation and publication

P1/P2/P3/P4 development line (local, unreleased): discovery/ implements the
patent-discovery skill (material intake → technical facts → patent candidates → human
confirmation → research handoff draft, G-I control layer). The drafting skill gained the P2
structured pipeline (DraftingPlan → ClaimAtom → Claim → SupportMapping → two-state writing
gate, G-D audit codes in drafting/references/14–15), the P3 professional-figure engine
(figures plan/render/validate, F-001…F-014 checks, drafting/references/16) and the P4
delivery pipeline (frozen PatentDocumentModel, DOCX/optional PDF, delivery manifest with
human sign-off, release_ready gate, CNIPA XML still not attempted; drafting/references/17)
plus the kernel-side patentkit draft / deliver commands. These are
included in the full package built from this tree; the paired kernel capabilities live in the
kernel source tree and only reach this repository through a rebuilt release/patches/ patch at
release-integration time. No release has been published for this line, and G-S / evidence /
validation semantics are unchanged.

Build a release package

python -m pip install ./patentkit-0.9.0-py3-none-any.whl
python tools/build_skill_release.py --out dist --built-at 2026-10-23T00:00:00+00:00 --json
python -m unittest discover -s tests -v

Requirements

• Python >= 3.11 (required by the skill scripts)
• Skill 3.8.0: patentkit >= 0.9.0, assessment protocol 1 (docx export needs the optional python-docx)
• Codex CLI >= 0.161 (the desktop app bundles a newer runtime, no upgrade needed)

Language

Since 2.5.0 the distributable skill runtime is Chinese-primary: SKILL.md, references/
(including the new 08–13 and glossary.md) and assets. The English edition is mirrored under
references.en/ (research 01–08 and drafting 00/09–13 maintained on both sides since 3.8.0). Script and
kernel messages, installers, and release notes remain English (engineering layer).
Implementation plans and verification notes under docs/ use Chinese.

Credits & license

The methodology was informed by ideas from these public Agent Skills
(no code or documentation is vendored):

• cn-patent-drafter (Apache-2.0)
• patent-disclosure-skill (Apache-2.0)
• patent-validity-checker (Apache-2.0)

This repository is licensed under the Apache-2.0 license (see LICENSE).

Security notes

• All bundled data is synthetic. Do not upload unpublished material to cloud clients.
• The patentkit kernel is distributed as a wheel; its source is maintained separately.

Optional extension: patent_intel (4.0 development line)

An independently versioned, strictly optional wheel (patent_intel, currently 0.1.0)
adds retrieval-recall tooling around the kernel without touching its authority:

pip install extensions/patent_intel            # stdlib-only base
pip install "patent_intel[kernel]"             # + read the kernel workspace
python -m patent_intel doctor --root <ws>      # read-only self-check
python research/scripts/recall.py plan --axis N1      # query-family proposals
python research/scripts/recall.py report              # offline recall diagnostic
python tools/score_retrieval.py --suite benchmarks/retrieval/fixtures/smoke --runs <runs> --json

Boundaries (see docs/adr/0001-kernel-boundary.md):
the kernel stays the sole authority for evidence levels, gates and sign-off; the
extension only produces proposals and read-only diagnostics inside its own sidecar
(<ws>/.patent-intel/, deletable without changing kernel validation). Online sources
(OPS) ship contract-tested but are not live-run from this repo; quota/limits
follow the provider portals. Architecture and data contracts:
docs/architecture/4.0-system.md,
docs/architecture/contracts.md.
Retrieval benchmark protocol: docs/retrieval-evaluation.md.

Version

Skill version: 3.8.0 (patent-research + patent-drafting; patent-discovery in the P1 development line),
companion kernel: 0.9.0. See CHANGELOG.md.