patent-application-creator

name	patent-application-creator
description	End-to-end patent campaign from ANY raw material ("here is some information; make a patent") to a filing-ready provisional package - invention mining, worth-it economics (design-around cost, detectability), exhaustive adversarial prior art, claims-first drafting, machine-verified compliance, hostile-examiner attack pass, and an honest no-go when nothing clears the bar or the fence is not worth the money
tools	Bash, Read, Write
model	sonnet

Patent Application Creator Skill

Run a complete patent campaign: take whatever the user has — a codebase, an invention disclosure, scattered notes — and either produce a filing-ready provisional package or a reasoned, evidence-backed explanation of why not.

This workflow was hardened by running it for real (a full campaign over a production codebase, July 2026). Every phase below exists because skipping it cost something in that campaign.

The contract

• The honest outcome is the deliverable. "Nothing here clears the novelty bar, and here is the prior art that kills each candidate" is a SUCCESS, not a failure. So is "this is patentable but not worth your money — file nothing, or publish defensively." Never inflate a weak candidate.
• The campaign guards the user's money, not just their filing. The population this tool serves cannot afford a technically perfect package for a commercially worthless patent. The worth-it question is asked BEFORE the expensive phases and re-asked every time the claims narrow — a campaign that only discovers "small fence" at the end has already spent the user's budget answering the wrong question.
• Machine checks gate, humans verify. Every automated finding marked LOW confidence gets manual verification; every clean bill of health names the checks that ran AND the checks that were skipped.
• Dates matter. Prior art is being published continuously; note recency threats found during the sweep and say so in the package. Filing sooner beats polishing longer for a provisional.
• The AI is never the inventor. The user is. The AI mines, searches, drafts, and verifies; the user's creation is what gets protected.

Phase 0 — Intake (any raw material)

Accept the rough ask as-is. If the material is a codebase, do NOT ask for a disclosure — mining is Phase 1's job. Ask only what cannot be derived: inventor name(s), and whether any of it has been publicly disclosed or sold (statutory bar dates). Then AUDIT the inventor's own public footprint — released products, demos, public repositories, marketing pages — and record first-disclosure dates: the inventor's own disclosures start the US grace-period clock and can immediately forfeit foreign rights. Do this before drafting, not after.

Phase 1 — Invention mining

Hunt CONCRETE TECHNICAL MECHANISMS, not features. For a codebase, fan out readers (one per subsystem) with this lens: a candidate must be (a) specific and implemented, (b) solving a technical problem, (c) arguably unconventional — never textbook auth/CRUD/caching. For each candidate capture: mechanism (how, not what), evidence location, problem solved, conventional alternative beaten, why the difference is non-obvious.

Then triage (kill/pursue) with these screens:

• Section 101 screen: billing, commerce, and organizing-human-activity subject matter draws Alice rejections regardless of cleverness — deprioritize.
• Crowded-field screen: thread pools, retries, distributed sync, testing methods — demand an unusual twist or kill.
• Unification screen: look for ONE principle several candidates instantiate; a system claim with multiple embodiments beats scattered small claims.

Phase 1-H — History mining (when the material is a git repository)

The current code shows WHAT exists; history shows what it BEAT, how hard it was, and when — and it holds mechanisms that no reader of HEAD can ever find, because they died in abandoned branches. For an inventor who rebuilds often, the abandoned branches are where the cleanest ground lives: a mechanism that never merged never deployed, so it has no disclosure clock running and full foreign rights intact. Run this pass alongside the Phase 1 code mining whenever the raw material has git history.

The funnel (validated on a real 808-commit, 9-branch + 92-PR-ref repo: ~2.5M agent tokens, minutes of wall clock):

1. Enumerate everything. Mirror-clone; fetch +refs/pull/*/head too. git rev-list --all is the universe. State the limit honestly: a host only serves commits still reachable from refs — force-pushed-away history is gone, so the stamp says "what the host still has," never "all that ever existed."
2. Collapse by patch-id. git rev-list --all | git diff-tree --stdin -p -r | git patch-id --stable — cherry-picks and rebases across hundreds of branches collapse to one entry each. Cost scales with unique patches, not branches. (Do NOT pass --no-commit-id; patch-id needs the commit lines to attribute patches.)
3. Mechanical triage into NAMED buckets. Exclude noise subjects (deps/docs/merges), lockfile/binary-only patches, and tiny diffs — every exclusion lands in a bucket with its subject recorded. Coverage must stay arithmetic: universe = read + excluded, reconcilable.
4. Read EVERY survivor. Batch patches (~30/agent, diffs capped ~7KB with the cap disclosed) and require each reader to return one verdict per patch AND the count of patches it saw — reconcile counts per batch. MECHANISM vs ROUTINE, with "what conventional approach the diff replaces" captured. Instruct readers explicitly: NEVER downgrade abandoned or unfinished work — abandoned cleverness is the target.
5. Cluster and locate. Same mechanism across many commits is one cluster. For each needle: git branch --contains and reachability against the default branch, then verify against the SHIPPED code whether the mechanism migrated (a bridge branch can make "stranded" commits' content live elsewhere — check the deployed source, not the branch topology). Only "absent from everything shipped" earns the no-disclosure-clock label.
6. Second adversarial pass until dry. One lens misses; re-read the ROUTINE pile with a "what did pass 1 misjudge?" prompt until a pass finds nothing new.

History also supplies what HEAD cannot: the diff shows the conventional alternative being replaced (Phase 1's hardest field to fill), a string of failed attempts before the breakthrough is objective non-obviousness evidence, and commit dates + authorship are conception records. Feed the surviving needles into the Phase 1 triage screens (including worth-it) like any other candidate.

Phase 2 — Prior art: every outlet, adversarially

Patents alone are NOT sufficient for software: the killing art usually lives in products, open source, standards, and papers. Sweep ALL of:

1. Patents: search_patents_bigquery (keyword search is term-AND, so 3-4 terms max), search_patents_by_cpc_bigquery (identify the CPC classes first), Google Patents pages for deep reads (free). Cost awareness: a default keyword search scans roughly 325 GiB (about $2 billed, or a third of a sandbox project's free month).
2. Non-patent literature: fan out adversarial web-research agents per claim cluster, each instructed to KILL the claim: products (what do the incumbent writing/coding assistants actually do?), open source (read the code), standards bodies, arXiv, engineering blogs.
3. Instruct every agent to return per-item verdicts: "what ours has that this lacks" or "ANTICIPATES". Aggregate into: CLEAN GROUND (recite in independent claims), KILL ZONES (never claim alone), MANDATORY CLAIM READS (references an attorney must read before the utility filing), and UNRESOLVED LEADS (bot-blocked pages and the like — escalate to the user, never silently drop).

Language discipline (non-negotiable). Sweep results support only statements of the form "no anticipation found in [outlets searched] as of [date]" backed by a dated element-by-element chart. Never write "nobody", "no art exists", "swept clean", or "survives" as facts — an agent's read of a repository is a snapshot of an evolving codebase, and a second read weeks later can contradict it. And a date rule that is easy to get wrong under urgency: published third-party art can never be outrun. Anything already public is prior art against any later filing, grace period or not; filing-urgency arguments apply only to disclosures that have not happened yet.

Decision gate: if no candidate has clean ground, write the no-go report (candidates, killing references, per-candidate reasoning) and STOP. That report is the deliverable. Then re-run the worth-it screen on each survivor AT ITS POST-SWEEP WIDTH: the clean ground is always narrower than the Phase 1 candidate, and the question is whether the SMALLER fence still costs a competitor anything. A survivor that is now avoidable by an obvious variant goes in the no-go report too — with its design-around named, and with the defensive-publication alternative stated (near-free, permanent, kills later third-party patents on the same mechanism).

Phase 3 — Claims-first drafting

Draft claims BEFORE the specification; the clean ground dictates them.

• Each independent claim's load-bearing limitation must be an element the sweep found nowhere. Elements found in isolation go in dependents.
• Proper form: "The method of claim N, wherein ..." — never shorthand (the analyzer's dependency tracking and the examiner both need it).
• Consistent terms: introduce with "a/an", reference with "the" using the IDENTICAL noun phrase.
• Wall off known art explicitly where cheap (for example "not by elapsed time" when the closest art expires suppression by time).
• Track the shrinking yard. Every wall-off and every narrowing amendment (here and again after the Phase 5 attack pass) makes the claim easier to design around. After the attack pass settles the final claim shape, re-run the worth-it screen one last time on what actually survived; record the verdict (design-around cost, detectability, provisional-vs-defensive-publication recommendation) in the package README so the user decides about the patent they can GET, not the one they imagined.

Then the specification: field, background (the problems, framed technically), summary (one paragraph per independent claim), brief description of drawings, detailed description covering EVERY embodiment with reference numerals, reduction to practice, and a variations paragraph (broaden: any model, any similarity metric, thresholds exemplary). The spec must contain each claim term VERBATIM — the support checker verifies this.

Source-verified claims (mandatory). Every limitation of every claim must be verified against the IMPLEMENTATION itself — open the code — not against a mining agent's summary and not against header comments, which go stale (a real campaign found a contract doc describing closest-match relocation while the executable body below it contained an ambiguity- refusal gate; the claim drafted from the summary was wrong in a way that contradicted the specification). The reduction to practice is the tiebreak for what claims, spec, and figures must all say — and they must all say the SAME algorithm, or each variant must be expressly a separate embodiment. Never claim behavior the implementation does not have (a "valid result of a preservation-type plan" reclassification that the code never performs is new matter waiting to be rejected).

Terminal-state completeness. Any specification asserting deterministic termination, bounded retries, or guaranteed outcomes must enumerate EVERY terminal state and every budget transition — including the unglamorous ones (mechanical exhaustion, global overrun backstops) — and the figure must show every branch with labeled edges.

Figures: create_block_diagram for the system (numbered components matching the spec), render_diagram with hand-written DOT per independent claim so decision branches carry Yes/No edge labels (create_flowchart does not label edges), with reference numerals on every node.

Phase 4 — Machine verification loop (iterate to clean)

1. patent-creator config set PATENT_ENABLE_ANTECEDENT_CHECK 1 — the antecedent check is opt-in; a campaign MUST run it.
2. review_patent_claims on the full claim set. Fix genuine findings (form, term drift); manually verify each remaining LOW-confidence flag and record the verdicts — that manual pass is part of the workflow, not optional.
3. review_specification (claims + spec) — iterate until ZERO critical issues: every flagged claim element gets woven into the spec verbatim.
4. check_formalities (abstract 50-150 words, title limits, sections).

Verification stamps carry content hashes, and edits invalidate them. Every recorded check result must state the SHA-256 (or equivalent) of the exact artifact text it checked. Any subsequent edit — adding one claim, touching one sentence — invalidates every stamp on that artifact; re-run the checks before packaging. A README that says "verified" about an artifact edited after its verification is the single most damaging falsehood a package can carry, because a reviewer who catches it stops trusting everything else. (This is the same input-fingerprint discipline the campaign's own subject matter implements for AI scans; apply it to yourself.)

Phase 5 — Two red teams, not one

5a. Hostile-examiner attack pass (claims). Spawn adversarial agents playing USPTO examiner against the final claims, armed with the sweep's closest references: strongest anticipation (102) argument, strongest obviousness (103) combination with per-limitation reference mapping and motivation to combine, and the Alice (101) attack plus its technical-solution counter. If an attack lands, narrow the claim or move the defeated element to a dependent — then re-run Phase 4 on the changed claims.

5b. Package red team (the whole artifact). Attacking the claims is not enough: a package can have perfect claims and still be unfit to file. First run the deterministic gate: check_package on the assembled directory. It recomputes every content-hash verification stamp (an edit after a recorded check is a critical failure), cross-checks claim counts against every "N claims" assertion, flags readiness language beside draft markers, catches commentary inside filing copies, and sanity-checks dates. Fix every critical before spending reviewer effort. Then hand the ASSEMBLED package — every file, nothing else — to a fresh adversarial reviewer with zero campaign context, tasked to find: internal contradictions between README, claims, spec, and figures; statements of verification that the artifacts themselves contradict; stamps that predate edits; date errors; strategy commentary that must not be filed; sweeping prior-art characterizations; and unsupported claim language. Nothing may be described as filing-ready until the package red team passes. (A real campaign's package — claims attacked and hardened — was caught by exactly this kind of fresh-eyes review with a stale "not yet compliance-checked" header beside a README saying "verified," a claim added after its recorded check, and three artifacts describing three different reconciliation algorithms. The claims red team caught none of that, because none of it was a claims problem.)

Phase 6 — Package

Working copies and filing copies are separate files, and the filing copies are generated mechanically. Filing documents contain claims and disclosure text only — zero bracketed strategy notes, zero checker scores, zero "vs art" commentary, zero next-steps sections, zero restriction strategy. Backgrounds make no categorical admissions about prior art ("known to the inventor" phrasing, not "no system does X"). Generate the filing copy by stripping the working copy with a script, then diff-check that nothing but claims/disclosure text remains.

A filing directory containing: specification and claims (filing copies), figures (converted to PDF meeting Patent Center requirements — letter/A4, embedded fonts, visually inspected after conversion; printing SVGs from a browser is not a QC process), the SB/16 cover sheet data (inventor residence, correspondence address), micro-entity certification if claimed (note: gross-income basis also requires small-entity eligibility, the counted-application limit, and qualification of ownership-interest holders), and a README with the pro-se provisional path and the 12-month utility clock, plus the mandatory-claim-read list and unresolved leads for the eventual attorney review. State plainly what the package is: a provisional built so the human legal review is fast — not a substitute for it — and state its verification status with content-hashed stamps, never with the word "verified" alone.

Failure modes this workflow exists to prevent

• Trusting a disclosure interview when the invention is in the code.
• Patents-only prior art (the killing reference is usually a product).
• Claiming an element the sweep found (it goes in a dependent, or dies).
• "100% compliant" reports with the load-bearing check silently off.
• Spec paraphrasing claim terms (support gaps surface in prosecution).
• Polishing past a publishing competitor (date pressure is real).
• Shipping a weak application instead of an honest no-go.
• A README that says "verified" about artifacts edited after the check (stamps without content hashes are lies waiting to be discovered).
• Claims drafted from an agent's summary instead of the source, or from a stale header comment instead of the executable body.
• Claims, specification, and figures describing three different versions of the same algorithm.
• Claim language describing behavior the implementation does not have.
• "Deterministic termination" asserted without enumerating every terminal state.
• "File before X" urgency about art that is already published (it cannot be outrun; only future art can).
• Strategy commentary, checker scores, and novelty absolutes left inside documents headed for the Patent Office.
• Attacking only the claims and never the assembled package.
• Spending the budget proving a claim patentable without ever asking whether the resulting fence is worth owning — worth-it is asked at triage, re-asked when the sweeps narrow the ground, and re-asked when the attack pass narrows the claims.