@diottodev/runic 1.0.0 → 1.1.0

package/research/patent/commands/cs-patent.md

@@ -0,0 +1,101 @@
+---
+name: "cs-patent"
+description: "/cs:patent <invention> — Patent prior-art + landscape intelligence with mandatory sub-use-case commitment. 6-Q grill-me intake (Q2 picks one of: novelty / FTO / landscape / diligence / litigation). Multi-source search (Google Patents + Espacenet + USPTO + optional Lens.org BYOK). 8-section .docx with verdict + claim text + family-resolved hits + mandatory legal disclaimer (novelty + FTO)."
+---
+
+# /cs:patent — Patent Prior-Art + Landscape Intelligence
+
+**Command:** `/cs:patent <invention description>`
+
+The `cs-patent` persona produces a sub-use-case-tailored patent dossier. **Refuses generic "patent help"** — must commit to one of 5 sub-use-cases at Q2.
+
+## Forcing Intake (6 Questions, One at a Time)
+
+| Q | Asks | Notes |
+|---|---|---|
+| Q1 | Invention (2-3 sentences, specific) | refuses vague; "AI for healthcare" pushed back |
+| Q2 | Sub-use-case: novelty / FTO / landscape / diligence / litigation | **Forcing — refuses "all of them"** |
+| Q3 | Jurisdictions (US/EP/CN/JP/KR/PCT/worldwide) | Asked only for FTO/landscape/diligence |
+| Q4 | Known prior art (patent number or paper) | Anchor; accept "none" |
+| Q5 | Risk tolerance: strict / signal-gathering | Asked for novelty + FTO |
+| Q6 | Attorney status (have you spoken to one?) | Asked for novelty + FTO; triggers disclaimer |
+
+Stop condition: after Q6 (or earlier with skips). Never re-open.
+
+## What You Get
+
+```
+patent_<invention-slug>_<sub-use-case>_<YYYY-MM-DD>.docx
+
+8 sections:
+1. Executive Summary + Verdict (NOVEL/CLEAR/FLAGGED/etc.) + legal disclaimer
+2. Closest Prior Art (5-10 ranked, claim-text extracted, hyperlinked)
+3. Patent Landscape (top filers, 10-yr trend, CPC distribution)
+4. Citation Graph Signals (foundational + recent high-cite, if Lens-enabled)
+5. Geographic Coverage (FTO/landscape/diligence only)
+6. FTO Flags (FTO only — risk per claim per jurisdiction)
+7. Strategy + Recommendations (sub-use-case-specific)
+8. Audit Log (searches, counts, plan-tier, attorney reminder)
+```
+
+## Per-Sub-Use-Case Behavior
+
+| Sub-use-case | Search emphasis | DOCX adjustment |
+|---|---|---|
+| Novelty | Narrow + claims-focused; pre-filing date irrelevant | Sections 5-6 abbreviated; verdict NOVEL/POTENTIALLY/NOT NOVEL |
+| FTO | Active patents only; jurisdiction-filtered | Section 6 expanded; verdict CLEAR/FLAGGED/HIGH RISK per jurisdiction |
+| Competitive landscape | Breadth + filer tally + CPC trends | Section 3 expanded; verdict = top-5 filers + 3 emerging entrants |
+| Acquisition diligence | Specific assignee + portfolio + assignment chain | Sections 3+5 expanded; ownership-verification flags |
+| Litigation prior-art | Target patent + adjacent art before priority date | Section 2 = ranked knock-out candidates |
+
+## Discipline
+
+- **Sub-use-case commitment mandatory** at Q2
+- **Sequential search 1 q/sec** across all sources
+- **CPC class follow-up** after initial keyword search
+- **Family resolution** — same-invention duplicates reported once
+- **Date discipline** — filing/priority/publication/grant distinguished
+- **Legal disclaimer mandatory** for novelty + FTO
+- **Source discipline** — only this session's tool calls
+- **Three-count tracking** — sent / received / cited
+- **Out-of-scope flagging** — trademark/copyright/trade-secret rejected at intake
+
+## Trigger Phrases
+
+- "prior art search for [invention]"
+- "patent search on [topic]"
+- "freedom to operate analysis"
+- "FTO for [product]"
+- "patent landscape for [field]"
+- "is [invention] novel"
+- "patents on [topic]"
+- "competitive patent analysis"
+- "prior art for litigation"
+- "patent diligence on [company]"
+
+## Anti-Patterns Rejected
+
+- Starting any search before user commits to a sub-use-case (refuses generic "patent help")
+- Batching all intake questions
+- Accepting vague invention descriptions
+- Keyword-only search without CPC/IPC class follow-up
+- Treating family members as separate hits
+- Confusing filing date with priority/publication/grant date
+- Skipping legal disclaimer when sub-use-case has legal consequences
+- Reporting verdict without claim-text evidence
+- Fabricating Lens.org citation data when key absent
+- Suggesting design-arounds without acknowledging attorney review required
+- Skipping audit log
+
+## Related
+
+- Agent: [`cs-patent`](../agents/cs-patent.md)
+- Skill: [`patent`](../skills/patent/SKILL.md)
+- Source spec: [`megaprompts/11-patent-megaprompt.md`](../../../megaprompts/11-patent-megaprompt.md)
+- Siblings: `/cs:litreview`, `/cs:grants`, `/cs:dossier`, `/cs:pulse`
+- Future: `/cs:syllabus`
+
+---
+
+**Version:** 1.0.0
+**Source:** Path-B direct conversion of `megaprompts/11-patent-megaprompt.md`

package/research/patent/skills/patent/SKILL.md

@@ -0,0 +1,287 @@
+---
+name: patent
+description: "Patent prior-art and landscape intelligence skill — not generic patent help. Commits to one of five sub-use-cases via forcing intake (novelty search / freedom-to-operate / competitive landscape / acquisition diligence / litigation prior-art) before any search runs. Searches Google Patents, Espacenet, USPTO, and optionally Lens.org for citation-graph signals. Output is an editable Word document (.docx) with verdict, ranked closest art (claim-text extracted), CPC-class-aware landscape, family-resolved hits, geographic coverage, FTO flags where applicable, strategy recommendations, and full audit log. Triggers: 'prior art search for [invention]', 'patent search on [topic]', 'freedom to operate analysis', 'FTO for [product]', 'patent landscape for [field]', 'is [invention] novel', 'patents on [topic]', 'competitive patent analysis', 'prior art for litigation', 'patent diligence on [company]'. Produces search signal, not legal advice — always recommends consulting a patent attorney before filing or licensing decisions. Trademark, copyright, and trade-secret questions are out of scope."
+license: MIT
+metadata:
+  source_spec: "megaprompts/11-patent-megaprompt.md"
+  build_pattern: "Path B (direct conversion)"
+  research_pack_convention: "Agent Integrity Rules verbatim per PR #657 audit; sub-use-case routing variant"
+  version: 1.0.0
+---
+
+# Patent — Prior-Art + Landscape Intelligence
+
+> **Portability:** Requires `web_fetch` (Google Patents, Espacenet, USPTO), `WebSearch` (adjacent academic art), Node.js with `docx` package, and optionally Lens.org API key for citation-graph signals. Works in Claude Code CLI natively. In Claude.ai with web tools + Code Execution + BYOK Lens.org, the workflow is supported.
+
+> **Out of scope:** trademark, copyright, trade-secret. These are flagged at intake. Use a different skill or qualified counsel.
+
+> **Legal disclaimer:** This skill produces search signal, not legal advice. Verdicts are technical assessments. **Always consult a patent attorney before filing or licensing decisions.**
+
+## Non-Generic Framing — The Differentiator
+
+This skill is **prior-art + landscape intelligence**. It **refuses to be a bucket**. Every invocation commits to one of five sub-use-cases via the grill-me intake before any search runs. The chosen sub-use-case dictates the entire search strategy, ranking heuristics, and DOCX emphasis.
+
+| Sub-use-case | Search strategy | DOCX emphasis |
+|---|---|---|
+| **Novelty search** | Narrow + claims-text focused; pre-filing date irrelevant | Closest art + claim-differentiation |
+| **Freedom-to-operate** | Broad + active patents only; jurisdiction-filtered | FTO flags + claim-by-claim risk |
+| **Competitive landscape** | Breadth + filer tally + CPC trends | Filer map + investment hotspots |
+| **Acquisition diligence** | Specific assignee + portfolio scope + assignment chain | Portfolio table + ownership verification |
+| **Litigation prior-art** | Specific target patent + adjacent art before priority date | Knock-out candidates ranked by relevance |
+
+See [`references/sub_use_case_routing.md`](references/sub_use_case_routing.md) for the canon.
+
+## Agent Integrity Rules (Research-Pack Convention)
+
+Locked verbatim per PR #657 audit.
+
+- **Execution discipline.** Sequential search calls only. **1 query/sec rate limit.** Confirm response received before next call.
+- **Source discipline.** Cite only patents returned by THIS session's tool calls. Training knowledge labeled `[Not from search — reference information]` and excluded from counts.
+- **Three-count tracking.** Queries sent / patents received (shown) / patents cited. Surfaced in audit log.
+- **Retry policy.** On failure → wait 3s → retry once → log. After 3 consecutive failures across tools: stop, alert user, explain what's missing.
+- **Plan-tier detection.** Lens.org free tier = 1000 queries/month. Google Patents has no auth but rate-limits per IP. Detect and surface caps.
+
+## Phase 1: Grill-Me Intake (6 forcing questions, one at a time)
+
+### Q1 (root) — Invention description
+
+> **Describe the invention in 2–3 sentences. What does it do, and what's new about it?**
+>
+> *Why I'm asking:* Concept and keyword extraction depends entirely on a precise description. Vague descriptions ("AI for healthcare", "a better widget") will be rejected — push back and ask the user to specify what the invention does and what differentiates it from existing approaches.
+
+**Refuse mush.** If answer is generic, ask once more: "What does it do that existing systems don't?" Then commit (with caveat in DOCX).
+
+### Q2 (depends on Q1) — Sub-use-case commitment
+
+> **What's the purpose of this search? Pick one:**
+>
+> 1. Novelty search (am I novel enough to file)
+> 2. Freedom-to-operate (will I get sued if I ship)
+> 3. Competitive landscape (who else plays here)
+> 4. Acquisition diligence (does target really own X)
+> 5. Litigation prior-art hunting (kill a specific patent)
+>
+> *Why I'm asking:* Each path uses a fundamentally different search strategy. I'll **refuse to start without you picking one**.
+
+Forcing format. If user says "all of them", push for the primary purpose — secondary purposes can run as follow-up searches.
+
+### Q3 (asked only if Q2 ∈ {FTO, landscape, diligence}) — Jurisdictions
+
+> **Which jurisdictions matter? Pick all that apply: US / EP / CN / JP / KR / PCT / worldwide.**
+>
+> *Why I'm asking:* FTO only matters where you'll sell. Landscape changes radically by region. Diligence requires checking all jurisdictions where the target operates.
+
+Skip for novelty (priority date is jurisdictionally portable) and litigation (jurisdiction is set by the target patent).
+
+### Q4 (depends on Q1) — Known prior art
+
+> **Have you already seen prior art close to this? Cite a patent number or paper.**
+>
+> *Why I'm asking:* If you know one piece of art, I can search adjacent to it — much more precise than starting cold. If you don't, that's fine — just confirm.
+
+Anchoring. Accept "none" but ask if the user has seen *any* related work even informally.
+
+### Q5 (depends on Q2) — Risk tolerance
+
+> **Risk tolerance for this search: strict (one close hit means abandon the path) or signal-gathering (you want the lay of the land regardless)?**
+>
+> *Why I'm asking:* Strict mode ranks aggressively and surfaces verdict-grade hits; signal mode prioritizes breadth and visualizations.
+
+Asked for novelty and FTO; skipped for pure landscape (always signal-gathering by definition).
+
+### Q6 (asked only if Q2 ∈ {novelty, FTO}) — Attorney status
+
+> **Have you spoken to a patent attorney? This skill produces search signal, not legal advice. Confirm you understand this is for technical assessment only.**
+>
+> *Why I'm asking:* Novelty and FTO have legal consequences. The skill's verdict is signal-grade; legal positions require qualified counsel.
+
+**Triggers the legal-disclaimer footer in the DOCX.** Skipped for landscape and diligence (lower legal exposure).
+
+**Stop condition:** After Q6 (or earlier if dependency skips applied), commit and start Phase 2. Never re-open intake after Phase 2 begins.
+
+## Phase 2: Search Strategy Selection
+
+Deterministic from intake answers. Use `scripts/sub_use_case_router.py`:
+
+```bash
+python ../scripts/sub_use_case_router.py \
+  --sub-use-case novelty \
+  --jurisdictions "" \
+  --risk strict \
+  --known-art "US10000000B2"
+```
+
+Returns: query plan (5-8 queries) + ranking heuristic + DOCX emphasis flags.
+
+## Phase 3: Multi-Source Search (Sequential)
+
+### Source priority
+
+1. **Google Patents** (https://patents.google.com) — workhorse, no auth required, broad coverage
+2. **Espacenet** (https://worldwide.espacenet.com) — global coverage, good for non-US art
+3. **USPTO PPS** (https://ppubs.uspto.gov) — US deep dive
+4. **Lens.org** (https://www.lens.org) — citation graph, BYOK API key required
+
+### Per-sub-use-case query patterns
+
+**Novelty:**
+- 3 narrow queries on invention-specific terminology (Google Patents)
+- 2 broad concept queries with synonyms (Google Patents + Espacenet)
+- 1 CPC-class-restricted query if class identified from initial hits
+
+**FTO:**
+- Jurisdiction-filtered: only active patents (not expired, not abandoned)
+- Date filter: priority < today
+- Active-claim text extraction for each hit
+
+**Competitive landscape:**
+- Broader queries on the technology space
+- CPC class identification → tally top filers in that class
+- 10-year filing trend by year per top-5 filer
+
+**Acquisition diligence:**
+- Specific assignee searches (target company + subsidiaries + named inventors)
+- Assignment chain check (USPTO assignment recordation)
+- Family resolution for deduplication
+
+**Litigation prior-art:**
+- Target patent input required (number)
+- Priority date extraction
+- Search for art before priority date in same CPC classes
+- Adjacent-claim-language search
+
+### Sequential discipline
+
+1 q/sec across ALL sources combined. Tracked via `scripts/citation_tracker.py` with timestamp-enforced gap.
+
+## Phase 4: Claim Extraction + Relevance Scoring
+
+For each closest-art hit:
+- Pull **independent claim 1** (the broadest claim — primary anticipation/obviousness vehicle)
+- Pull **key dependent claims** (claims that add the inventive step)
+- Score relevance against invention description (overlap of claim language with Q1 terminology)
+
+Rank by score. Verdict per sub-use-case (NOVEL / POTENTIALLY NOVEL / NOT NOVEL for novelty; CLEAR / FLAGGED / HIGH RISK per jurisdiction for FTO).
+
+## Phase 5: Citation Graph + Family Resolution
+
+### Citation graph (Lens.org BYOK)
+
+If user provides Lens.org API key:
+- Foundational-patent identification (cited-by count > threshold, typically 50+)
+- Recent high-cite signals (citations in last 24 months as proxy for current activity)
+- Forward citations from target patent (litigation prior-art) or from closest art (novelty)
+
+If no Lens.org key: skip; note in audit log; recommend manual citation review on Google Patents.
+
+### Family resolution
+
+Same invention often filed in multiple jurisdictions (US + EP + JP + CN). Group by family ID or priority number to avoid double-counting. Use `scripts/family_resolver.py`:
+
+```bash
+python ../scripts/family_resolver.py --hits-file hits.json
+# Returns: deduplicated family list + family-member jurisdictions
+```
+
+## CPC/IPC Classification Awareness
+
+**Critical:** keyword search alone misses adjacent art. After initial search, extract the CPC/IPC classes from top 5 hits and run **one class-restricted query**. This consistently surfaces art that keyword search misses.
+
+See [`references/cpc_classification_canon.md`](references/cpc_classification_canon.md) for the canon.
+
+## Phase 6: DOCX Generation (8 Sections)
+
+Sub-use-case-dependent emphasis. Via Node.js + `docx` library.
+
+1. **Executive Summary + Verdict** — Sub-use-case banner + one-line verdict (NOVEL / FLAGGED / etc.) + 3-4 key findings + legal disclaimer footer
+2. **Closest Prior Art** — 5-10 patents in ranked order. Per hit: hyperlinked title + assignee + filing/priority dates + independent claim 1 text (italicized) + relevance score + relevance rationale (1-2 sentences)
+3. **Patent Landscape** — Top filers table (top 10 by count) + 10-year filing trend description + CPC class distribution table. Only for landscape and diligence; abbreviated otherwise.
+4. **Citation Graph Signals** — Foundational patents (if Lens-enabled) + recent high-cite activity. If Lens unavailable, note "manual review recommended" and skip table.
+5. **Geographic Coverage** — Filings by jurisdiction for top 10 hits. Only for FTO, landscape, diligence; skipped for novelty and litigation.
+6. **FTO Flags** (FTO only) — Active patents posing infringement risk. Per flag: hyperlinked patent + jurisdiction + relevant claims + risk level (HIGH/MEDIUM/LOW) + mitigation note.
+7. **Strategy + Recommendations** — Sub-use-case-specific:
+   - Novelty → claim differentiation suggestions
+   - FTO → design-around hints + jurisdiction strategy
+   - Landscape → who-to-watch list
+   - Diligence → red flags in portfolio
+   - Litigation → ranked knock-out candidates
+   - **Mandatory disclaimer to consult patent attorney** for any filing/licensing decision.
+8. **Audit Log** — Searches table (#, query, source, results, status), counts (sent/shown/cited), tool constraints (plan-tier notes), failed steps, attorney-consultation reminder
+
+### Styling
+
+Arial 12pt body, navy headings (#1a3a5c), light blue table headers (#e8f0f8), red FTO-flag callout. `ExternalHyperlink` patterns:
+- Google Patents: `https://patents.google.com/patent/[number]`
+- Espacenet: `https://worldwide.espacenet.com/patent/...`
+- USPTO: `https://patents.uspto.gov/patent/...`
+
+## Date Discipline
+
+Distinguish at every hit:
+- **Filing date** — when the application was first submitted
+- **Priority date** — earliest claim of priority (often earlier than filing)
+- **Publication date** — when the application became public (typically 18 months after priority)
+- **Grant date** — when the patent was granted (later than publication)
+
+Surface the **legally-relevant date** per sub-use-case:
+- Novelty → priority date (vs invention's anticipated filing date)
+- FTO → grant date + status (active vs expired)
+- Landscape → publication date (when public knowledge began)
+- Diligence → grant date + assignment date
+- Litigation → priority date of target patent (sets the prior-art cutoff)
+
+## Phase 7: Deliver
+
+- Save: `<output-dir>/patent_<invention-slug>_<sub-use-case>_<YYYY-MM-DD>.docx`
+- Chat summary: file path + sub-use-case + verdict + audit counts + plan-tier
+- Validate: `python scripts/office/validate.py <docx>`
+- Reminder: "Consult patent attorney before filing/licensing"
+
+## Tooling
+
+| Script | Role |
+|---|---|
+| `scripts/citation_tracker.py` | Multi-source three-count audit (Google Patents + Espacenet + USPTO + Lens.org) at `~/.patent_sessions/<session>.json` |
+| `scripts/family_resolver.py` | Group same-invention filings across jurisdictions by family ID / priority number |
+| `scripts/sub_use_case_router.py` | Deterministic search-strategy selection from intake answers |
+
+## References
+
+- [`references/sub_use_case_routing.md`](references/sub_use_case_routing.md) — 5-sub-use-case canon (7+ sources)
+- [`references/cpc_classification_canon.md`](references/cpc_classification_canon.md) — CPC/IPC class follow-up rationale (7+ sources)
+- [`references/legal_disclaimer_discipline.md`](references/legal_disclaimer_discipline.md) — when + why disclaimer mandatory (7+ sources)
+
+## Error Handling
+
+| Failure | Behavior |
+|---|---|
+| User refuses to commit to sub-use-case | Refuse to proceed. Re-ask Q2 with examples. |
+| Invention description is generic | Reject answer. Re-ask Q1 with "what does it do that existing systems don't?" |
+| Google Patents rate-limits | Wait 3s, retry once. Fall back to Espacenet for that query. Log in audit. |
+| Lens.org key missing | Skip citation graph section, note "manual review recommended" in DOCX. |
+| Claim text extraction fails | Fall back to abstract; flag as "abstract-only" in relevance rationale. |
+| Family resolution incomplete | Note in audit; same-invention duplicates may appear; suggest manual deduplication. |
+| All searches return <3 hits | Surface explicitly as "either niche art or genuine gap"; never fabricate. |
+| 3 consecutive tool failures | Stop, alert user, explain what's missing. |
+| DOCX generation fails | Save raw data as JSON fallback so user doesn't lose work. |
+| Target patent number invalid (litigation) | Validate format before search; ask user to confirm. |
+
+## Anti-Patterns To Reject
+
+- Starting any search before user commits to a sub-use-case (refuses generic "patent help")
+- Batching all intake questions instead of one at a time
+- Accepting vague invention descriptions ("AI for healthcare")
+- Keyword-only search without CPC/IPC class follow-up
+- Treating family members as separate hits (must be deduplicated)
+- Confusing filing date with priority date with publication date
+- Skipping the legal disclaimer when sub-use-case has legal consequences
+- Reporting a verdict without claim-text evidence
+- Fabricating Lens.org citation data when key is absent
+- Suggesting design-arounds without acknowledging attorney review is required
+- Skipping the audit log
+
+---
+
+**Version:** 1.0.0
+**Source spec:** [`megaprompts/11-patent-megaprompt.md`](../../../../megaprompts/11-patent-megaprompt.md)
+**Build pattern:** Path B (direct conversion). Research-pack sibling, sub-use-case routing variant.

package/research/patent/skills/patent/references/cpc_classification_canon.md

@@ -0,0 +1,154 @@
+# CPC/IPC Classification — Why the Class Follow-Up Catches What Keywords Miss
+
+This reference answers exactly one decision: **why does the patent skill always run a CPC/IPC class-restricted query after initial keyword searches, and how does class follow-up systematically surface art that pure keyword search misses?**
+
+## The Core Claim
+
+Keyword search alone systematically misses adjacent prior art. Patent attorneys describe this as the **"different vocabulary problem"**:
+
+- A 1995 patent on "machine learning for image recognition" might describe its invention as "neural network for visual classification" — different terminology, same underlying concept
+- A patent in semiconductors might use "transistor channel" where a software paper would use "data flow path" — same idea, different field's vocabulary
+- A patent might intentionally use unusual terminology to **broaden claim scope** (legal strategy)
+
+The CPC/IPC classification system was designed precisely to bridge these vocabulary gaps. **Always-run class follow-up** is the skill's mechanical correction for keyword-only blind spots.
+
+## What Are CPC and IPC?
+
+| System | Maintained by | Granularity | Used by |
+|---|---|---|---|
+| **CPC** (Cooperative Patent Classification) | USPTO + EPO | ~250,000 classes | All major patent offices since 2013 |
+| **IPC** (International Patent Classification) | WIPO | ~75,000 classes | Used as fallback in some jurisdictions |
+
+Both are hierarchical. Examples:
+
+- `G06N` → Computer systems based on specific computational models (CPC + IPC)
+- `G06N3/00` → Computing arrangements based on biological models
+- `G06N3/04` → Architecture, e.g. interconnection topology
+- `G06N3/045` → Combinations of networks (deep learning hidden layers)
+
+When a patent examiner classifies a patent, they assign one or more CPC classes. Patents in the same class are conceptually related EVEN IF they use different vocabulary.
+
+## The CPC Class Follow-Up Pattern
+
+After initial keyword search returns top 5 hits:
+
+1. **Extract CPC classes** from those 5 hits
+2. **Tally** to find the dominant class (1-3 classes typically)
+3. **Run one class-restricted query** — same keywords + CPC class filter
+4. **Compare** results to initial keyword-only results
+
+**Empirical observation:** the class-restricted query consistently surfaces 2-5 additional hits that the keyword search missed. Some of these are highly relevant (the "vocabulary mismatch" cases).
+
+## Concrete Examples
+
+### Example 1: AI for medical diagnosis
+
+| Search | Top results |
+|---|---|
+| Keyword "AI medical diagnosis" | 2017+ patents using "AI" / "ML" / "deep learning" + "diagnosis" |
+| **+ CPC G16H50/20** (medical informatics for diagnosis) | + 1990s patents on "expert systems" + "decision support" — same concept, different era's vocabulary |
+
+Without the class follow-up, the searcher would miss 25 years of foundational expert-system art that the USPTO clearly considers prior art.
+
+### Example 2: 3D printing materials
+
+| Search | Top results |
+|---|---|
+| Keyword "3D printing polymer" | 2010+ patents using "additive manufacturing" + "polymer" |
+| **+ CPC B33Y70/00** (materials for additive manufacturing) | + 1980s patents on "stereolithography resin" — predecessor terminology |
+
+### Example 3: Recommender systems
+
+| Search | Top results |
+|---|---|
+| Keyword "recommendation algorithm" | 2010+ patents using "recommender" / "collaborative filtering" |
+| **+ CPC G06Q30/0631** (recommender system for products/services) | + 1990s patents on "preference matching" + "user modeling" |
+
+## Why This Matters Per Sub-Use-Case
+
+### Novelty
+
+Missing class-adjacent art = **false negative**. User concludes invention is novel; later examiner finds the missed art and rejects. Class follow-up prevents this expensive surprise.
+
+### FTO
+
+Missing class-adjacent active patents = **false confidence**. User ships product believing it's clear; gets sued by patent owner whose patent used different vocabulary. Class follow-up surfaces these.
+
+### Litigation prior-art
+
+Missing class-adjacent art before priority date = **weak invalidity case**. The art that would knock out the target patent might be using completely different vocabulary; class follow-up finds it.
+
+### Landscape + Diligence
+
+Class follow-up surfaces the **technology lineage** — which classes the field operates in, who files in each class, how the field has evolved.
+
+## Operational Pattern
+
+In `Phase 3` of patent's SKILL.md:
+
+```
+1. Run initial keyword queries (per sub-use-case)
+2. Extract CPC classes from top 5 hits → identify 1-3 dominant classes
+3. Run ONE class-restricted query: keywords + CPC class filter
+4. Merge results, deduplicate, rank
+5. (Optional) If multi-class: run one query per dominant class
+```
+
+The class follow-up is a **single additional query per dominant class** — minimal budget cost, high signal yield.
+
+## How to Identify the Right Class
+
+After initial search, look at the top 3-5 hits' classification fields. Most patent search interfaces (Google Patents, Espacenet, USPTO PPS) show CPC classes in the metadata.
+
+**Heuristic:**
+- If 3+ of top 5 hits share a class → that's your dominant class
+- If hits are spread across many classes → dominant class is the most-frequent across the top 10 hits
+- If still spread → run class follow-up for top 2 classes (2 extra queries)
+
+## Anti-Patterns
+
+### Skipping class follow-up "to save queries"
+
+The skill's query budget per sub-use-case explicitly allocates 1-2 queries for class follow-up. **Skipping it to save 1 query is the most common false-economy** in patent search. The signal-per-query of class follow-up consistently exceeds keyword-only queries.
+
+### Relying solely on top-1 hit's class
+
+The top-1 hit might be an outlier. Look at the top 3-5 hits' shared classes for the dominant class.
+
+### Treating IPC and CPC as interchangeable
+
+CPC is more granular and modern (post-2013). When available, prefer CPC classes. Fall back to IPC for older patents that haven't been re-classified.
+
+### Class follow-up without CPC class identification
+
+Just running "G06N" with no further specificity is too broad. Use the most specific class that 2+ top hits share (e.g., `G06N3/045` not just `G06N`).
+
+### Ignoring class signals in DOCX
+
+The dominant CPC classes ARE valuable signal for the DOCX. Surface them in Section 3 (Patent Landscape) so the user understands the technology classification of their search space.
+
+## Operational Checklist
+
+- [ ] Initial keyword queries run (per sub-use-case)
+- [ ] CPC classes extracted from top 5 hits
+- [ ] Dominant class(es) identified (1-3)
+- [ ] Class-restricted query run (additional 1-2 queries)
+- [ ] Results merged + deduplicated + ranked
+- [ ] Dominant CPC classes surfaced in DOCX Section 3 (or Section 2 for novelty)
+- [ ] Audit log notes class follow-up as part of search strategy
+
+## Citations (7 sources)
+
+1. **CPC Scheme — USPTO + EPO joint maintenance.** https://www.cooperativepatentclassification.org. Authoritative source for the CPC hierarchy + per-class definitions. The skill recommends consulting CPC scheme for any class beyond top-3 frequency.
+
+2. **WIPO IPC Strategic Plan + IPC Schema.** https://www.wipo.int/classifications/ipc. Source for IPC fallback discipline (used for pre-2013 patents that lack CPC reclassification).
+
+3. **Mowery, D. C., Nelson, R. R., Sampat, B. N., & Ziedonis, A. A., *Ivory Tower and Industrial Innovation* (Stanford U Press, 2004).** Source for the historical analysis of how patent classifications evolve over time and why cross-era keyword search fails. Empirical evidence for the "different vocabulary problem".
+
+4. **WIPO PATENTSCOPE search documentation.** https://patentscope.wipo.int. Source for cross-jurisdictional class search syntax (especially for non-US/EP jurisdictions).
+
+5. **Cohen, W. M., Nelson, R. R., & Walsh, J. P., "Protecting Their Intellectual Assets" — *NBER Working Paper* 7552 (2000).** Source for the empirical evidence that keyword-only patent search systematically under-reports prior art (especially in fast-moving technology areas).
+
+6. **MPEP §901 — *Manual of Patent Examining Procedure* (USPTO).** Source for the examiner-side discipline of using CPC classes for prior-art search. The skill mirrors examiner discipline by including class follow-up as mandatory.
+
+7. **Lemley, M. A., & Sampat, B., "Examiner Characteristics and Patent Office Outcomes" — *Review of Economics and Statistics* 94(3), 2012.** Source for the empirical analysis showing experienced examiners use CPC classes more aggressively + produce stronger prior-art rejections. Class follow-up is the experienced-examiner technique.