Patent Application Creator

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