# 交底 Desk (Jiaodi Desk)

> A Chinese patent disclosure workbench at https://jiaodi-desk.skillsafe.ai/ —
> paste one project's technical material and take it to a hand-off-ready 技术交底书
> (Chinese patent technical disclosure) in one sitting: mine the patent points,
> draft the disclosure from the desensitized template, plan the CNIPA novelty
> search, then self-check logic closure and formula-parameter consistency.
> Primary UI language: Simplified Chinese. Chinese patent practice only.

## What it does

Four lanes over the same work object (the project's material and its evolving
disclosure draft), run by an AI model behind a metered API:

- **mine（专利点挖掘）** — extracts patentable points from pasted project
  material. Each point is a complete problem → technical means → beneficial
  effect triple with a novelty first-read (high / medium / low), a
  prior-art-risk note describing which conventional practice it may collide
  with (never a concrete patent number), and the missing information a drafter
  would need. Recommends which point to draft first.
- **draft（交底书成文）** — writes the full 技术交底书 from the material and a
  chosen point, following the standard desensitized template: 发明名称, 技术领域,
  背景技术 (with the required defect statement), 发明内容 (technical problem /
  solution / effects), 附图说明, 具体实施方式, 术语表. Figures are emitted as
  mermaid source; the page renders them in-browser (a built-in flowchart-subset
  renderer) with per-figure PNG download. Company names, emails, codenames are
  replaced with placeholders and each replacement is registered. Sections the
  material cannot support are marked 【待补】 rather than fabricated.
- **novelty（查新检索方案）** — produces the patent-search PLAN for a human to
  execute: distinguishing technical features, CN/EN keyword matrices with
  synonyms, IPC main-group candidates, pasteable query strings for CNIPA /
  Google Patents / Espacenet, per-feature knockout criteria and an execution
  protocol. The model is prohibited from naming any patent number. The page can
  additionally EXECUTE a non-patent-literature search server-side (declared
  providers web.wikipedia and web.duckduckgo via POST /v1/app-api/search):
  results come back provenance-stamped (record_id, url, retrieved_at), the user
  curates them, and a re-run returns body.comparison — a per-feature
  hit/miss/unclear table over exactly the kept records, with citations verified
  against the sent list. A miss is not novelty; the patent-database search
  remains human-executed.
- **check（交底书自检）** — the source skill's signature pre-hand-off self-check
  (accepts the same curated search_results and flags draft novelty claims that
  collide with a retrieved record):
  §8.4 logic closure (problem, solution and effects must interlock; broken
  links are listed), §8.5 formula-parameter consistency (every symbol defined,
  units coherent, ranges consistent with the embodiment), section-by-section
  review, desensitization re-check, figure-reference consistency, and rewritten
  passages for every serious defect.

## The free lane

A deterministic scanner (jdscan.js) runs entirely in the browser before any
sign-in: desensitization risks (company-style names, emails, reachable
addresses, phone numbers, project codenames, secrecy markings), template
section coverage, background-defect statement, figure references vs the figure
list, formula symbols used but never defined, undefined acronyms, absolute
wording (最优/唯一), vague parameters (约/若干), thin embodiment sections, and
whether claimed effects carry any quantification. Every flag is sent with the
paid run and the model must reconcile each one (confirmed / set-aside /
superseded); unreconciled flags are surfaced to the user.

## Intake and export

Material can be pasted, or loaded as multiple files / a whole project folder
(docs and code; prose files are concatenated first, code after, all read
locally via FileReader - nothing uploads until a run). Results export as
Markdown, JSON, findings CSV, and Word (.docx) built entirely in the browser -
the .docx embeds the figures as PNGs rendered from the mermaid source.

## Contract

Input: JSON with `task` (mine|draft|novelty|check), `patent_type_hint`
(auto|invention|utility|design), `material`, `disclosure`, `point`, `context`,
optional `search_results` (kept records from /v1/app-api/search — the model may
cite only their record_ids), and the scanner's `prescan` flags. Output: one JSON envelope with lane,
invention, title, patent_type, posture (ready|needs-work|blocked), verdict,
summary, assumptions, open_questions, findings (id/severity/area/evidence/why/
fix/fix_text), coverage_check, artifact (the disclosure or plan as Markdown),
next_lane, and a per-lane body. Full worked examples per lane:
https://jiaodi-desk.skillsafe.ai/api.html

## Sources and boundaries

Derived work based on the open-source skill @handsomestwei/patent-disclosure-skill
(https://github.com/handsomestwei/patent-disclosure-skill, MIT) — its Mode A
disclosure-writing pipeline (挖点 → 成文 → 查新 → 自检) adapted into a web
workbench. Not affiliated with the skill's author, CNIPA, or any patent agency.
Output is working-draft material for a licensed patent practitioner to review:
not legal advice, no grant predictions, no fee or deadline guidance, and no
search is ever performed. Chinese patent practice only. The model works only
from pasted material and is prohibited from inventing facts, data, patent
numbers or literature.

## Pages

- https://jiaodi-desk.skillsafe.ai/ — the app
- https://jiaodi-desk.skillsafe.ai/api.html — API tutorial (cURL, Python,
  JavaScript, Go, Java, Ruby, PHP, C#), input fields and output contract