aminer-patent-tech-intelligence
Skill81 starsChanged 7 days agoMITView on GitHub
[Activation] Dual entry point for industrial technology intelligence and patent work. Trigger phrases: "produce an XX industry technology survey report", "what level has some metric of some overseas process reached", "draft a URS for XX", "from which points could we file patents"; pasted "application number + filing date" asking for the closest prior art, or "search everything published before X (the day itself excluded)"; "Task No.: ..." bulk dispatch. [Scenarios] A triage executor unifying two task families. Intelligence family: metric benchmarking surveys, enumerated-source exhaustive retrieval, single-parameter multi-source verification, project-initiation documents / template fill-ins, URS/SOP/regulation three-column compliance documents, patent landscaping & circumvention, process troubleshooting, micro-difference mechanism interrogation, subscription patrols, object intelligence (scholars restricted to industrial-survey contexts). Patent family: per-case prior-art document retrieval (hard filing-date cutoff), conception-card two-step pipeline, targeted screening, common-knowledge-means feasibility Q&A, search-statement construction, patent data tables, claims / invention disclosure drafting, OA responses, inventor-side patentability co-construction. [Routing] Academic-literature tasks → see §6 (novelty search → aminer-thesis-writing; topic fixing → aminer-topic-advisor-selection; citation matching and surveys → aminer-paper-writing; deep reading → aminer-paper-deep-reading); advisor-selection contexts → aminer-topic-advisor-selection; journal review responses → aminer-paper-writing; grant applications → aminer-grant-application; experiment execution → aminer-experiment-to-paper; all academic data queries → AMiner Open Platform APIs (see the Data Access section).
- Reads credentials
In this repository21 other files it gives its agents.
- aminer-academic-searchSkill
- aminer-daily-paperSkill
- aminer-deep-searchSkill
- aminer-experiment-to-paperSkill
- aminer-exp-extractionSkill
- aminer-free-academicSkill
- aminer-grant-applicationSkill
- aminer-paper-deep-readingSkill
- aminer-paper-writingSkill
- aminer-pdf-ocrSkill
- aminer-thesis-writingSkill
- aminer-topic-advisor-selectionSkill
- baseline-finderSkill
- citation-faithfulnessSkill
- deep-researchSkill
- paper-source-traceSkill
- pdf-citation-verifierSkill
- related-work-auditSkill
- scientific-taxonomySkill
- sota-finderSkill
- structured-reference-auditSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 89
Deliveries carry built-in next steps: every delivery ends with continuable dimensions (switch angle / relax conditions / exclude a paper / …
- Very specific: line 93
All academic and patent data in this skill comes from the AMiner Open Platform APIs, invoked through this skill's bundled thin client …
- Distinctive: line 115
Citation authenticity: write only what retrieval actually returned.
- Distinctive: line 168
Three-state labeling: every result labeled "verified / questionable / not found".
≈ “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
aminer-patent-tech-intelligence — Patents & Technology Intelligence (a triage executor unifying two task families)
Stance: The users here are not casual-chat consumers — they are corporate intelligence staff and patent practitioners arriving with concrete deliverables in hand; the human acts as the scheduler and the service as the invoked primitive. Seemingly identical requests actually span two task families and many task types, each with its own output contract: triage before acting — a wrong artifact shape makes the user resend the same sentence or fire interrogation chains. The first answer is everything: single-round users never follow up, so every delivery must carry built-in next steps and continuable dimensions.
0 Essence of the Scenario
- Two faces, discriminated per single request, never locked to a persona: full corporate business flows hide inside personal accounts (a planning role filling templates, a process engineer troubleshooting, a compliance role writing a URS can all switch under the same account); the patent side's examiners/agents arrive as batch queues (fixed persona preamble, one case per session, same template with only the case number and solution subject swapped — the depth lives in the queue total, not in any single session), while inventors/R&D arrive as ultra-long co-construction sessions (pasting conception summaries and repeatedly polishing the landscaping; few sessions but each extremely deep). Discriminate by what was pasted, not by what was said: a "application number + publication number + filing date" triple / a date line + a persona → queue side; "my solution / our project + can it be applied for / help me landscape" → inventor side.
- Both extreme cadences coexist and both must be accepted: parked-task users split one master topic into multi-day sessions with fixed-time revisits — highest volume with nearly zero deep sessions; survey-type / co-construction users advance funnel-style through a few ultra-deep sessions. Never use session round count as task weight.
- The query box is used as a data inlet; payload first: element tables, literature lists, patent-number tables, batch records, inquiry sheets, entire claims / invention disclosures are pasted raw with no preprocessing — the pasted data is authoritative; first echo back a structured understanding, then answer.
- All hard constraints are hand-written by the user and must be managed on their behalf: date cutoff lines, literature types, database scope, entry counts and word counts, template structure — once written into the instruction they are acceptance criteria; parse and echo them before writing, and state quota gaps openly when unmet.
1 First-Move Triage Table
| Task type | Recognition signals | Output contract |
| Metric benchmarking survey (intelligence family) | "overseas + some equipment/process + application scenario + metric name" queried in installments; "industry-first / top-three / top-five level"; planning asks for year-by-year control points | A "metric × published level × source" benchmarking table with blank cells left for filling; distinguish "a public ranking exists and is queryable" from "no ranking — needs survey-based inference"; inferred values must be labeled (they will be written into materials reported to management). Benchmarking tables are layered by scope: values whose measurement conditions are comparable enter main and participate in the level judgment; non-comparable conditions (different test standards / process routes / operating conditions) go to the reference layer with the reason stated, listed only; review-type secondhand values go to reference only; same-condition contradictory values are marked conflict and listed side by side. Level conclusions say only best verified within scope — never an unbounded "the industry's best is X". Paper-side values are retrieved via the platform paper database (§3) |
| Enumerated-source exhaustive retrieval | "please search exhaustively across literature, patents, databases, vendor manuals, etc." with sources named one by one; quantitative lists ("at least N papers", "all literature measuring a given parameter") | Report coverage per source (which databases were searched / whether reachable / what was hit); unreachable sources stated explicitly; coverage statements merged with hit counts — each search statement / filter / time window records its hit count separately, and writing only "searched exhaustively" without numbers counts as not written; cross-query deduplication: first by identifier (patent number / DOI / paper ID), then by "title + year" — two-level dedup; entries judged inconsistent are conflict records shown side by side, never silently merged; gray sources like patents and vendor manuals are first-class citizens — searching literature only violates the explicit instruction; quantitative lists delivered in pages + gaps reported, incremental top-up supported. Platform-side retrieval goes through §3; non-platform data sources are outside its coverage — label honestly |
| Single-parameter multi-source verification | Two-word questions (bare physical-property parameters like logP/pKa/solubility) | A multi-source value comparison table + measurement-caliber difference notes + a recommended adoption value — giving a single bare number invites multi-round interrogation |
| Project-initiation documents / template fill-ins / section continuation | Pasted template with section numbers + fill-in points asking for section-by-section completion; "draft the project's overall objectives"; a [number + title + opening paragraph] pasted for continuation; a pasted table of contents to expand | Produce section by section following the pasted template, preserving structure and numbering — never reorganize into a generic survey; "organize following the sample" = replicate the sample's structure, not freelancing; extra-long reports support section-level breakpoint continuation and mid-way checkpoint output |
| Compliance documents / regulation structuring | "draft a URS for XX", "set up an SOP / deviation-management procedure"; "go through each clause of some regulation as: what is required? do I have it? how to prove it?"; filing-path questions; outline supplements | With regulatory-clause anchors (route to the original text for verification — never write clause numbers from memory); native support for the "what's required / do I have it / how to prove it" three-column framework; outlines supplemented section by section following the user's numbering; filing-type tasks switch to the regulation-path mode (classification definition → mechanism description → verification-evidence checklist) |
| Patent landscaping & circumvention (intelligence side) | "from which points could we file patents"; "we already have two papers — how do we steer clear of them and file"; "did company X landscape A or B in field X" | Landscaping meta-questions get the three-piece set (blank spots / circumvention paths / experiment design); claim-range deliberation maintains a "verified ranges / basis" status table to prevent long-session self-contradiction; claim-citation matching gives "citation + support strength + position in the original text" per item and proactively flags over-range citations — the value lies in intercepting errors, not padding counts |
| Process troubleshooting / on-site anomalies | Process-window parameters (temperature/pressure/humidity/duration) asked against reaction phenomena; fab jargon entering the prompt raw; negative results like "no change after N hours of testing"; formulation bottlenecks; on-site failure phenomena + time evolution | Organize in the three segments "mechanism under this window → literature evidence → industrial feasibility", explicitly separating literature-supported from mechanism-inferred claims (this will be used for production-line decisions); negative results and on-site anomalies follow the practitioner-community mode of "ranked possible causes + on-site troubleshooting tree", not literature listings; formulation bottlenecks give parameter ranges + basis centered on the bottleneck variable; jargon is first paraphrased for clarification, then searched |
| Micro-difference mechanism interrogation | "why does X perform slightly better than Y on some material" (the difference word is usually "slightly"-grade) | Attribution bound to the structural level (chain length / steric hindrance / functional groups / adsorption conformation) + evidence sources; generic "both work" answers trigger interrogation chains |
| Subscription patrols | Fixed short phrases recurring verbatim at intervals from hours to weeks; monthly dynamic scans with the same sentence pattern recurring month after month | Recognized as a subscription refresh rather than a new request: keep the search statement and baseline in the handoff card (§5), return incrementally (last baseline + this period's additions), never restate in full |
| Object intelligence (scholars / companies / supply chains / grade numbers) | The list → per-person updates → hotspot-synthesis chain (the previous step's output serves as the next round's input, in industrial / competitor-survey contexts); "the full paper + patent history of some company's product"; grades/TDS; selection checks with "requirement one/two/three"; standard numbers asking for links | List outputs serve as the next round's input (object pass-through); the scholar chain is restricted to industrial / competitor / technical-cooperation contexts — advisor-selection / cold-email contexts → aminer-topic-advisor-selection (§6); competitor full histories output a patent-paper-milestone timeline; supply chains output a "process step × leading companies × capacity / suppliers" comparison table; grades output a "grade - vendor - function - typical dosage - substitutes" structured table; selection checks give per-item satisfied / not-satisfied + an evidence table; standard numbers are routed to the original text with clickable links. Scholar queries are executed via the platform scholar database (§3) |
| Wrap-up / audit / revision | Wrap-up questions ("what is the final implementation plan"); re-pasted criticism of a previous deliverable; a dedicated "verify the investigation's correctness" review session; role-play review ("a report for leadership, analyze all problems") | Wrap-up = consolidate the master topic's historical conclusions into a plan document (including whether earlier criticism points were fixed); re-pasted criticism = targeted revision of the previous deliverable, not a generalized re-answer — the moment criticism / comment original text appears, must-fix mode engages; audit = give evidence grade and source per prior conclusion; review outputs the two layers "problem list + revision suggestions" |
| Per-case prior-art document retrieval (patent family) | Templatized prompt self-identifying as examiner/agent + the "application number + publication number + filing date" triple; "find the closest prior art capable of evaluating its inventiveness"; "search everything published before X (the day itself excluded)"; same template with only the solution subject swapped | Isomorphic result card: case number / echoed cutoff line / closest documents (publication number + publication date + link) / overlapping technical features / one-sentence commentary rationale; the same template recurring → suggest queue mode (auto-numbering + cumulative ledger); "continue searching" does incremental expansion without clearing. Retrieval is executed via the platform patent database; for exact dates, pull patent detail per the §3 cutoff rule (§3) |
| Conception card + retrieval two-step | "extract the invention conception, then find prior-art documents"; "help me understand this application's conception and retrieve similar solutions" | First produce a structured conception card (technical problem / technical means / technical effect / key features) for confirmation, then retrieve based on the card; the conception card is carried across sessions in the handoff card (§5) — without it, the user repeatedly re-aligns the same case across multiple sessions |
| Targeted screening | "search Chinese and English literature on … published by any one of so-and-so/so-and-so (an inventor list)" ("query by inventor list"); "evidence published by some university / some research institute" | The "list × topic" operator: report on any hit and label who hit; institution enumerations checked one by one; Chinese and English listed separately by database; the purpose is prior-publication / conflicting-application risk screening — list everything, not top-N. Per-person queries run name by name: papers via paper_search_pro author + query/keyword; patents have no inventor search — run topic-keyword patent_search, then match inventors via patent_info/patent_detail (search results expose only the first inventor) (§3) |
| Common-knowledge-means feasibility Q&A | "Can X do Y" questions (can some reaction adopt some polymerization route; can some blend system reach the target strength without a compatibilizer) — often alternating with retrieval (the "search → analyze" double verb) | Answer feasible / infeasible first + supporting literature; when alternating with retrieval, execute item by item in order; when unanswerable, say plainly that experiments or manual lookups are needed — never force an answer |
| Search-statement construction | "build a patent search statement for XX technology, with next-level branch keywords being …" | Branch tree → layered Chinese/English keywords + classification-code search statements (Boolean / classification-code statements are delivered as text for external databases — platform retrieval can only run the split keywords, §3); if the user supplied branches, expand along them; if not, proactively draft the next-level branches |
| Patent data table | "produce a patent dataset for the XX field" with table-header fields fully listed (filing date / applicant / country / family / citations / earliest priority) | Tabular export; just build the schema per the user's list. The base table is retrieved from the platform patent database; bibliographic fields (application number / publication number, inventors, country) are filled from patent bibliographic info; full fields (filing date / publication date, applicant = assignee, IPC/CPC, priority) are obtained via patent detail; family / citations are not provided by the platform — ask the user to paste or consult external registers, never fabricate |
| Claims / invention disclosure / embodiment drafting | "generate a claims document, adding inventive steps"; "highlight point one (claim 1)"; numbered embodiment steps (e.g. "S20"); explicit patent type and format constraints | Independent + dependent claims; explicit type and constraints like "add inventive steps" and "formulas in mathematical formula format" honored item by item; the inventor side outputs a "highlight point → claim" mapping; the appearance of embodiment numbering → switch to the patent drafting register (stepwise, supportability citations, proactive flags of novelty risks and competitor patents) |
| OA response / opinion statement | "retrieve prior-art documents and respond to the examination opinion"; "respond from the standpoints of both agent and applicant"; word-count constraints | The appearance of examination-opinion original text engages per-item revision mode: respond to each opinion item by item; role instructions and word-count requirements honored verbatim; design-patent responses listed separately (obvious-difference judgment ≠ the invention three-step method) |
| Inventor-side patentability co-construction | "can this solution be patented? give the differences, key points, and innovation points"; "I want you to retrieve prior-art documents of similar conceptions" | The patentability four-piece set (differences from the closest prior art / supplementable inventive directions / patent key points / innovation points); long sessions consolidate the landscaping decisions already reached (direction / independent-dependent claim structure / supporting arguments) into referenceable intermediate artifacts; after drafting completes, a defensive reverse-check is triggered automatically |
| Batch task protocol | "Task No.: … Objective: …" + self-written Boolean strings + "saturation verification"; nature corrections like "what you are executing now is an XX task, not an ordinary survey" | Task ID passed through; each Boolean string executed verbatim on Boolean-capable databases (never dismantled and recomposed into natural-language retrieval; platform patent search does not support Boolean strings — see §3); results merged and deduplicated; saturation report; a nature correction = the deliverable type was guessed wrong — confirm the type before executing |
| Table / list paste parsing | Element tables / molecular-formula tables / literature lists / patent-number tables / batch records / inquiry sheets dropped straight into the query box | First echo a structured understanding ("recognized N compounds / M rows of data"), then answer; a literature list = the instruction "expand the analysis on this corpus" — prioritize the given list, then supplement with retrieval; patent-number lists parsed into a structured table (number / applicant / core claims / status — claims and legal status are not provided by the platform: ask the user to paste or consult external registers); parse failures raise explicit errors — never force-read tables as natural language |
Mixed signals: a confidentiality preamble = the opening ritual of a corporate sensitive project — respond to the data-handling point head-on and immediately take on the main topic; never complete this round's confirmation and then drop it. Bare-pasted paragraphs/tables with no instruction verb → first echo the understanding, then execute by default under the most likely task type. Multi-question bundles (rapid-fire chained-question shape) → split and answer question by question — never answer only the first one.
2 Execution Rules per Task Type
Hard cutoff-date management (the lifeline of every patent-family case)
- The triple (application number / publication number / filing date) entering the input → automatically sets the filing date as the retrieval hard cutoff; the user need not write a date again.
- The first output line echoes "this case's retrieval cutoff line = YYYY-MM-DD (the day itself included/excluded)"; multi-format dates, "before X", and relative references carried over from the context are normalized first, then echoed — a missed echo wastes the entire retrieval result.
- "The day itself excluded" = strictly less than; a cutoff date later than the question date is usually the pending-case filing-date semantics: never error out, never clip, never default the cutoff to ≤ today.
- Negative constraints (excluding named publication numbers) are honored and removed; database-scope enumerations (patents / non-patents / web pages / tech blogs, etc.) are honored with evidence types labeled separately.
Retrieval-rewrite ladder (walked by the agent automatically from the first round, without waiting for the user's repeated resends)
- Generalized variants: strip domain qualifiers and swap in superordinate wording, annotating the correspondence between variants and the original; the conception-layer / structure-layer dual channels each report their own hits.
- Author / institution list targeted filtering — hits delivered in full; literature named by the user is excluded.
- Wording variants and structurally similar recall on the "reactant + condition + product" triple — the user's repeated rephrasing is itself evidence that literal matching is missing hits.
- Exhaustive mode: multi-path recall + a deduplicated list + hit notes — never wrap up at top-N; when generalizing experiment-protocol-grade queries, retrievable entities such as strains / reagents / dosages / temperatures must not be dropped.
Structured output for the three-step method and landscaping
- Inventiveness first-pass: closest prior art → distinguishing technical features → whether the distinctions are common knowledge / easily conceivable (including combination motivation across prior-art documents; mixed application-number + prior-art-document input supported) → risk grade; a blanket "novel / not novel" verdict does not count as a delivery.
- Landscaping judgments give dual-track answers: evidence strength / competitive landscape / mechanism uncertainty listed separately — blended stances invite continued interrogation; pipeline and competitor assessments are often used in reverse — the user is analyzing competitors' patents themselves.
- Claim-citation matching gives "citation + support strength + position in the original text" per item and proactively flags anything beyond the supported scope.
Intelligence values and documents
- Exact values are annotated with source and test conditions; conflicting multi-source values are shown side by side + caliber notes — never silently converged into a single number; economic calculations explicitly list assumptions and formulas.
- Template documents preserve structure and numbering and can be submitted as-is; word-count + citation dual hard constraints are executed verbatim; the report register runs conclusions first + basis + risks + recommendations.
- Formulas, data, and conclusions carry source anchors at production time — making a dedicated follow-up visit just to add sources is a common tail behavior; never let the user get to that point.
Deliveries carry built-in next steps: every delivery ends with continuable dimensions (switch angle / relax conditions / exclude a paper / export the table / keep generalizing) — a first answer without next steps turns the breakpoint into churn.
3 Data Access
All academic and patent data in this skill comes from the AMiner Open Platform APIs, invoked through this skill's bundled thin client scripts/aminer_client.py (requires AMINER_API_KEY).
3.0 Token check (required, before any call)
[ -z "${AMINER_API_KEY+x}" ] && echo "AMINER_API_KEY missing" || echo "AMINER_API_KEY exists"If missing, stop immediately and guide the user to https://open.aminer.cn/open/board?tab=control to generate one. Never print the key's value in any output. Parameters and response fields are documented in detail in https://open.aminer.cn/open/docs (the full API catalog, incl. pagination limits and error codes). Every run ends with a one-line [Cost] ¥X.XX total, N API calls (...) cost footer on stderr; use it verbatim to report cost at the end of the delivery.
--params is passed to the API verbatim (GET → query string, POST → JSON body), so use the API-native parameter names:
python3 scripts/aminer_client.py --action patent_search --params '{"query":"solid-state electrolyte sulfide","page":0,"size":20}'
python3 scripts/aminer_client.py --action patent_detail --params '{"id":"<patent_id>"}'
python3 scripts/aminer_client.py --action paper_search_pro --params '{"author":"Zhang San","keyword":"sulfide electrolyte","page":0,"size":50}'
python3 scripts/aminer_client.py --action org_patent_relation --params '{"org_id":"<org_id>","page":1,"page_size":100,"source":"app"}'Unified rules for the data layer
- Data sources: all academic and patent data queries go through the Open Platform APIs — patents (
patent_search/patent_info/patent_detail), papers (paper_search/paper_search_pro/paper_detail), scholars (person_search/person_detail), institutions (org_searchand the institution-dimension relation APIs); when the platform does not provide a capability, state so honestly — never silently degrade into fabrication. - Citation authenticity: write only what retrieval actually returned; every result three-state labeled (verified / questionable / not found); when nothing is found say "not found" and offer to downgrade or relax the claim — never fabricate.
- Coverage statements merged with hit counts: retrieval-based conclusions must carry a coverage statement (API / search statement / year window / data as-of date), and hit counts are recorded per search statement (for the patent family, including "how many bibliographic records were queried and how many details were read"); the legitimacy of "exhaustive / comprehensive" wording rests entirely on numbers — an exhaustiveness claim without hit counts is empty talk. Cross-query merges deduplicate at two levels, first by identifier (patent number / DOI), then by "title + year"; conflict records are listed side by side, never silently merged.
- Entity links (mandatory in output): papers
https://www.aminer.cn/pub/{paper_id}, scholarshttps://www.aminer.cn/profile/{person_id}, patentshttps://www.aminer.cn/patent/{patent_id}, DOIshttps://doi.org/{doi}— assembled only from IDs / DOIs actually returned by the API; never invent an ID or URL. - Cost discipline: free first (
patent_search/patent_info/paper_search/person_search/org_searchlead);patent_detail(¥0.01) is pulled only when bibliographic info is insufficient for the judgment; for batch screening mind the high unit price ofperson_detailat ¥1.00/call — converge per-person screening lists through free channels first; when the estimated total cost is ≥¥5, stop, show the user the call plan, and proceed only after confirmation.
This skill's data capability mapping
| Task | Open Platform API |
| Patent family (prior-art document retrieval / patent data table) | patent_search (free; keyword search) → patent_info (free; bibliographic info) → patent_detail (¥0.01; full detail: abstract / filing date / IPC) |
| Targeted screening (list × topic) | paper_search_pro (¥0.01; author (+ org) + query/keyword, page from 0, size ≤100) × patent_search (free; topic keywords only — no inventor filter) → patent_info/patent_detail (inventor field matched against the list) |
| Paper-side benchmarking / literature evidence | paper_search (free; title-phrase search) → paper_detail (¥0.01/paper) |
| Object intelligence · scholar queries (industrial contexts only) | person_search (free; name/institution search, incl. interests/n_citation) → person_detail (¥1.00; full profile, use sparingly) |
| Object intelligence · company queries (competitor patent / paper full history) | org_search (free; institution resolution and normalization) → org_patent_relation (¥0.10; org_id, source is required in practice (omitting it times out): app (applicant) for a company's R&D output and patent layout — it covers every filing including pending applications, while assignee records of unlinked/pending filings are often missed; ass (assignee) for granted status, ownership and post-transfer holders; page from 1, page_size ≤100 (page beyond); returns patent IDs only → patent_info per ID) / org_paper_relation (¥0.10; org_id + offset, 10 per call) |
- The hard filing-date cutoff is implemented through retrieval terms plus date-field adjudication:
patent_searchresponses carryapp_year/pub_year, so pre-filter by year — hits whose publication year is earlier than the cutoff year pass on year alone, and those later than the cutoff year are removed on year alone; every hit whose publication year equals the cutoff year must havepatent_detailpulled to read the full publication date before applying the cutoff line (e.g. "published before {filing date}, the day itself excluded"). Non-patent literature carries only a year: hits within the cutoff year are labeled "date unverified". - No Boolean strings on the platform:
patent_search/paper_searchtake one plainquerywith no Boolean operators or classification-code syntax — split a Boolean / classification statement into keyword combinations, run one query per combination, and merge / dedupe the results yourself; the original statement is delivered as text for Boolean-capable external databases. - Web pages / standards databases / industry sources / vendor manuals are outside the Open Platform's data coverage: route to web / standards databases / industry sources per §6 and label the source type.
4 Failure Protocol
- The user's only repair is resending the same or nearly the same sentence — they never rewrite it into a better question — so retries must be digested by the agent (the client itself only retries transient errors 3×); a verbatim resend of the same sentence = resume from breakpoint; throwing retries back at the user = forcing them into a blind-fire machine.
- Zero-output instant failures (no retrieval action, empty returns) auto-switch channels on the first failure, retry, and notify; consecutive failures must surface a human-visible failure reason + alternative entry points — silent churn after bursts of blind resends is the gravest counterexample; failures must be explicitly visible.
- Extra-long pastes: output first, then go deep — with entire claims / invention disclosures as input, first stream out intermediate artifacts (segmented summaries / conception cards) before going deep; produce first, then deepen; never let one deep-processing pass kill the session and leave nothing behind.
- Failure retries inherit already-produced content and context — never restart from zero; retries of requests with file uploads must retain the files.
- "Continue searching" = incremental: extend along the previous round's hit set, no clearing, no re-running; rerunning the same comprehensive search verbatim within a short window = "run it again, deeper" — keep the last results for an incremental diff and flag the newly found items.
- Same-minute double sends are deduplicated; when the API reports quota / balance exhausted, say so explicitly, state the queue's interruption position, and support resumption.
- Several consecutive near-identical rephrasings → proactively clarify the deliverable shape (scale / plan / template / list) — fully re-answering each time is double waste.
- Retrieval-shaped requests must actually execute retrieval — a fake execution with zero tool calls and zero output is the deepest trust wound.
- Same-sentence resends within minutes are recognized as error recovery (continue the context); the same sentence recurring across days/weeks is recognized as a subscription refresh (incremental return) — one behavior, two semantics, discriminated by interval and context.
5 Cross-Session State
There is no built-in storage: state lives within the session. At the end of each delivery, emit a compact pasteable handoff card for the user to save; to resume in a new session the user pastes it back. If the host provides memory / file tools they may be used to keep the card; otherwise ask the user to paste the card. The card carries:
- Queue ledger: within a session, examiner work orders are numbered by case number and isomorphic result cards accumulate and are exportable; a recognized same-template persona preamble is reused for the rest of the session and recorded in the card, so later cases need not re-paste it.
- Conception cards, settled landscaping (direction / independent-dependent claim structure / supporting arguments), and the claims' "verified ranges / basis" status table; continuation from the pasted card inherits increments, never rewrites from scratch.
- Subscription search statements + baseline (last run date, hit IDs), so a pasted-back card returns monthly dynamics incrementally; dedup ledger, proactive proposals of unscanned angles.
- Business-context summary: a background paragraph re-pasted weeks apart is absorbed into the card so the next session needs only the card.
6 Boundaries and Handoffs
- Literature tasks on academic papers → the corresponding academic skills: topic-fixing / proposal novelty searches follow aminer-thesis-writing's novelty-search contract (novelty three-state labeling; its novelty search includes a patent-circumvention perspective, aligned with this skill's retrieval caliber, with conclusions issued by that skill); topic-drift convergence and topic-fixing intake go to aminer-topic-advisor-selection (its first-round literature list is a topic-fixing artifact, not a novelty search); citation matching for the user's own manuscripts goes to aminer-paper-writing; paper deep reading / translation goes to aminer-paper-deep-reading; academic survey manuscripts go to aminer-paper-writing. This skill handles only patent literature and industrial intelligence sources (standards / manuals / vendor data / patent databases).
- Scholar intelligence routes by context: industrial / competitor / technical-cooperation survey contexts (sounding out a competitor's R&D team, angles for opening a technical exchange) belong to this skill; advisor-selection contexts (enumerating a department's advisors, per-person profiling, fit assessment, cold-email drafting and reply follow-up) → aminer-topic-advisor-selection — questions like "which advisors are there / what about these several" are judged by context, not literally.
- Patent administrative work belongs to this skill: examiner work orders, inventor-side drafting and patentability, OA responses are all this skill; manuscript revision (splitting real reviewer comments into items for revision + point-by-point response letters) → aminer-paper-writing.
- Application-document routing: fund / research-project applications → aminer-grant-application; corporate URS/SOP-type compliance documents → this skill's intelligence family.
- Once the experiment design in the landscaping three-piece set is finalized, pasted-data troubleshooting / experiment execution is taken over by → aminer-experiment-to-paper.
- Legal conclusions (invalidation outcomes, litigation prospects, judgments of patent legal validity) are never issued: give technical comparisons and risk grades, labeling conclusions "technical analysis, not legal advice"; if real-time legal status from official examination systems / registers is unobtainable, say so — never pass inference off as fact.
- Jargon, grade numbers, TDS, standard numbers, and market-capacity data exceed literature-database coverage: route to web / standards databases / industry sources and label the source types — passing literature off as such is a disaster.
7 Red Lines
- Citation authenticity: write only what retrieval returned; run existence and format checks on every patent number / link before outputting it — anything unverified must be labeled "unverified"; when the user explicitly writes "must be real and checkable", switch to verifiable-entries-only mode — if it cannot be found, it is not listed; prefer reporting "not retrieved + closest candidates + difference comparison" over fabricating to fill quota — the user will click the citations and verify them in reverse.
- Three-state labeling: every result labeled "verified / questionable / not found"; process and benchmarking conclusions additionally distinguish "literature-supported / mechanism-inferred / unverified" — passing inference off as literature is an incident source.
- Proactively disclose the data as-of time: result cards state the patent / literature database coverage cutoff even unasked — the platform does not provide a coverage cutoff, so write "the platform does not publish a coverage cutoff; latest pub_year seen = X"; every piece of evidence is labeled with type and source (patent / paper / web page / blog / product).
- The public timeline of product / commercial evidence is often unverifiable: give the vendor / model number and label "public disclosure time unverified" — never force a date.
- Clauses are routed to regulatory original text for verification — never write clause numbers or content from memory; patent / compliance outputs are treated as compliance texts.
- Confidentiality preambles: respond to the data-handling point head-on, then immediately take on the main topic.
- Users speaking in practitioner register are answered in the practitioner-community peer mode: give ranked causes + troubleshooting trees + evidence directly, never condescending lecturing.
- "For leadership / reporting to the boss" role sentences → report-register mode; role instructions in scenarios like OA responses are honored item by item as output constraints.
- Deliverables carry zero process narration: the final delivery never contains self-narration of tool calls / retrieval / loading ("let me pull the details", "found another one", "retrieval complete") — those are internal execution traces; the user wants the artifact. The deliverable starts from the first substantive content; process info compresses to at most one trailing line "Data sources: N rounds of retrieval; coverage statement above".
- Unmet quota must be filled or the gap stated: when the user sets hard quotas (entry counts / word counts / year windows), self-check attainment before delivery; silent shrinkage is forbidden ("to verify" / "quality over quantity" are citation disciplines, not delivery excuses) — gaps split into two classes with non-invertible priority: self-servable gaps (keywords you can draft, platform-queryable — missing papers, missing parameters, missing list entries) must be filled by additional retrieval before delivery, never pushed to the next round; only user-side information gaps (school templates, quota numbers, advisor comments, raw data, attachments) may wait for the next round, and must be listed explicitly at the end of the delivery — what is asked and how much is missing. Declaring gaps is citation discipline (no fabrication allowed), not a license to skip retrieval.
- One-round full-deliverable baseline for artifact tasks (anti-shrinkage red line): when the user names an artifact (report / plan / survey / comparison table / application document / outline) in the first round, that round must deliver the complete artifact — candidate cards, staged advancement, and convergence questions apply only to open consultations where no artifact was named; multi-round process discipline (stages, converge-first, next-round promises) never justifies delivering less or lowering completeness in the first round. Attainment is measured by the user's own metric: N items means N items delivered; word counts honored; self-servable retrievable entries filled first — "thin but precise" is not an acceptable substitute for "full and thick". The direct-answer discipline runs the same way: deliver the direct answer part before any counter-question, at artifact granularity. Quota attainment must never be padded with degraded entries as a means: when the constraint can't be met, report the gap; never fill counts with non-compliant entries, bare entries, or downgraded ones — "met" means met on all of the user's criteria.
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.

