Patent Search skill for Codex

A useful search is reproducible and honest about its limits: the invention is broken into elements, every query and database is logged, the closest references are mapped element by element in a claim chart, and the output is handed to a patent attorney or agent, who makes any legal call.

This is research support, not legal advice. Patentability, validity, infringement and freedom-to-operate opinions require a registered patent attorney or agent. Many companies also ask engineers not to search or read third-party patents without counsel’s direction, because knowledge of a patent can affect damages in an infringement case. Ask the user whether their counsel has approved the search before starting an infringement-related one.

Pick the search type

Goal	Search type	Date cut-off	Focus
Can we patent this?	Novelty / patentability	Before the planned filing date	Any public disclosure worldwide, any language
Is that patent valid?	Invalidity / prior art	Before the target patent’s priority date	Claims of the target patent
Can we ship this?	Freedom to operate (FTO)	In-force patents only	Claims of live patents in the markets you sell in; counsel-led
What exists in this field?	Landscape	Usually last 10-20 years	Assignees, trends, white space

Workflow

- [ ] 1. Confirm purpose, counsel involvement and confidentiality
- [ ] 2. Write the invention as numbered elements
- [ ] 3. Build vocabulary: synonyms, jargon, older terms, CPC codes
- [ ] 4. Search patent databases; log each query
- [ ] 5. Search non-patent literature
- [ ] 6. Screen, then read claims of the closest hits
- [ ] 7. Build the claim chart; check status and families
- [ ] 8. Deliver the search report with limits

1. Confidentiality

Do not paste an unfiled invention’s details into public tools, forums or AI services outside the user’s control; public disclosure can destroy novelty in most countries. Use generic queries, and keep detailed notes in the user’s private workspace. Searching needs network access.

2. Elements

Invention: on-device detection of keyboard typing cadence to unlock a laptop
E1 capturing inter-key timing intervals during password entry
E2 building a per-user timing model on the device
E3 comparing a new entry's timing against the model
E4 granting access only when both password and timing match
E5 updating the model after each successful login

3. Vocabulary and classification

• Synonyms and older terms: “keystroke dynamics”, “typing rhythm”, “keystroke biometrics”, “dwell time”, “flight time”. Patents use deliberately broad language (“input device”, “temporal characteristics”).
• Find CPC codes by looking at the classifications on the 3-5 most relevant patents you find with keywords, then browse the CPC scheme around them. Searching by CPC catches documents that use unexpected wording.

4. Databases

Database	Strength
Google Patents	Fast full-text, many offices, machine translations, “similar documents”, includes some non-patent literature
Espacenet (EPO)	~150M documents, strong classification search, INPADOC family and legal status
USPTO Patent Public Search	Official US full text (replaced PatFT/AppFT), advanced field searching
WIPO PATENTSCOPE	PCT applications, cross-lingual search
The Lens	Free, links patents to scholarly works, exportable result sets

Search pattern: start with element-combination keyword queries, then CPC + a few keywords, then citation chasing (backward citations of the best hits, and documents that cite them). Log every query:

| # | Database | Query / filters | Date run | Hits | Reviewed | Relevant |
|---|---|---|---|---|---|---|
| 1 | Google Patents | ("keystroke dynamics" OR "typing rhythm") AND (unlock OR login) AND "on device"; priority before 2026-01-01 | 2026-10-28 | 412 | 60 | 5 |
| 2 | Espacenet | CPC G06F21/32 AND typing | 2026-10-28 | 230 | 40 | 3 |

Exact operator syntax varies by database and changes; use each database’s own advanced search form and help page rather than assuming one syntax works everywhere.

5. Non-patent literature

Prior art includes anything publicly available before the relevant date: papers (Google Scholar, IEEE Xplore, ACM DL, arXiv), product manuals, standards, open-source repositories (commit dates), conference talks, archived web pages (Wayback Machine captures give dates). For software, NPL is often where the closest art is.

6. Reading patents

• The claims define legal scope; the description and abstract do not. Read independent claims (those not referring to another claim) first.
• Kind codes: US A1 is a published application (not granted); US B1/B2 are granted patents; EP A1/A2 are applications, EP B1 is granted. Applications may never grant or may grant with narrower claims.
• Dates: priority date (earliest claimed), filing date, publication date, grant date. Applications publish about 18 months after priority, so the most recent ~18 months of filings are invisible to any search.
• Families: the same invention filed in several countries; review one member and note the others (Espacenet “also published as”, INPADOC family).

7. Claim chart and status

| Element | Ref A: US 1,234,567 B2 (prio 2014-03-02) | Ref B: Smith et al., 2012 paper | Ref C: open-source repo X (commit 2016-05) |
|---|---|---|---|
| E1 timing during password entry | Yes - col. 4, l. 10-25 | Yes - sec. 3.1 | Yes - capture.py |
| E2 on-device model | Partly - server-side model, fig. 2 | No | Yes |
| E3 compare | Yes - claim 1 | Yes | Yes |
| E4 password AND timing | Yes - claim 3 | No | No |
| E5 update after login | No | Yes - sec. 4.2 | No |

Report what each reference discloses with pinpoint citations (column/line, paragraph, figure, section). Leave conclusions about anticipation or obviousness to counsel; you can note that no single reference covers all elements, or that a combination appears to.

For in-force questions, check legal status (Espacenet INPADOC legal status, USPTO Patent Center): granted, pending, lapsed for unpaid fees, expired, or withdrawn. Term is generally 20 years from the earliest non-provisional filing date, with adjustments that only counsel should rely on.

Gotchas

• US law has a one-year grace period for an inventor’s own disclosure; most other jurisdictions require absolute novelty. “We’ll blog first and file later” can forfeit rights outside the US.
• Machine translations of foreign patents are good enough to screen, not to rely on; flag key foreign references for professional translation.
• A long list of loosely related hits is less useful than three references charted well.
• Patent titles and abstracts are often vague by design; screen on claims and figures.
• Assignee names vary (subsidiaries, acquisitions, misspellings); search variants for landscapes.
• Never state that something “is patentable”, “is not infringing” or “is invalid”.

Verification

• Every query in the log can be rerun and gives a similar hit count.
• Each charted reference has publication or priority dates before the relevant cut-off.
• Pinpoint citations open to the stated passage.
• The report lists databases not searched, languages not covered and the 18-month gap.

Report template

# Prior-art search: <invention short name> (<date>) - for counsel review
**Purpose:** novelty search before filing | **Cut-off date:** <date>
**Elements searched:** E1-E5 (see appendix)

## Closest references
1. <Ref A> - discloses E1, E3, E4; differs on E2 (server-side), E5 absent
2. ...

## Claim chart
## Search log (databases, queries, dates, hits)
## Limits
Not searched: <databases, languages>; filings from the last ~18 months unpublished;
NPL coverage limited to <sources>.
## Questions for the attorney