gfodor/legal-skills: Markdown Agent Skills for U.S. Utility-Patent Work

The billable process that legal-skills is aimed at

The README frames patent work as a document pipeline: facts that are not in the draft, a specification and claims that either enable the invention or do not, a search that either found the dead-ringer or did not, an examiner who rejects under 35 U.S.C. 101, 102, 103 or 112 until the claims are allowable, and a competitor who keeps the value proposition while dropping the limitation you treated as essential. The claim is that firms wrap that pipeline in scarcity, letterhead and a rate card, and that the wrapper is what the repository removes.

The intended user is narrow. The README names inventors, founders and in-house engineers, and says the goal is to displace legal fees rather than assist counsel. It also draws a line: lawyers who work on judgment, strategy and courtroom work keep their work, while lawyers whose product is walking a known statute and filling known forms are the target. That is a positioning statement, not a feature list, and it should be read as one.

Three skills, run in order, each with a SKILL.md contract

The repository holds three directories: patent-audit, patent-examine and patent-workaround, plus LICENSE and README.md. Each is described as a complete agent skill: a SKILL.md the orchestrator reads as its contract, with payloads, templates and scripts that keep the model honest. The pipeline is sequential. patent-audit answers whether the draft is even fileable and stops on statutory bars, producing ranked defects and a parsed.json. patent-examine runs a recurrent simulated USPTO examination and amends only after a simulated Office Action. patent-workaround runs adversarial design-around testing.

The iteration rule is the interesting part. The README states you iterate examine against workaround until two examiners and two adversaries clear the same packet, and that a prior audit is not a substitute for examination, nor a prior examination a substitute for workaround testing. patent-examine and patent-workaround will use patent-audit's parser and mechanical checks if that sibling skill is installed. The audit itself is described as 316 checklist items with a deterministic pre-pass, and the claims work as a 64-item claims part plus a mechanical parser that builds the dependency graph. The stated reason for the scripts is blunt: models cannot count, cannot do date arithmetic, and will report all numerals consistent after finding 19 of 20.

Installing the skills and running a first audit

The README describes these as agent skills: a directory with a SKILL.md the model is instructed to follow, pointing at the Anthropic agent-skills documentation. The install section in the README begins by saying to copy the three skill directories, and the README text is truncated at that point, so the exact destination path is not documented in what is available. The repository layout is the reliable reference: clone it and you have patent-audit/, patent-examine/ and patent-workaround/ as top-level directories.

git clone https://github.com/gfodor/legal-skills.git
cd legal-skills
ls patent-audit patent-examine patent-workaround

Each of those three directories is expected to contain a SKILL.md. If your agent loads skills from a fixed location, copy the directories there rather than pointing at the clone, because the sibling lookup between patent-examine and patent-audit depends on the skill being installed, not merely present on disk.

A first real use follows the pipeline order. Feed inventor facts and the draft into patent-audit and expect ranked defects plus parsed.json; stop if the audit surfaces a statutory bar, since the README says the audit stops on those. Only then move to patent-examine, and only after a simulated Office Action should claims be amended. The README does not document a CLI entry point, flags, ports or environment variables for any of the three skills, so there is nothing to invent here: the interface is the SKILL.md contract the orchestrator reads.

What the skills refuse to do, and where they can hurt you

The disclaimer is long and specific, and it is the most useful part of the repository for anyone deciding whether to adopt it. The project states it is not a law firm, creates no attorney-client relationship, and produces no patentability, infringement or freedom-to-operate opinion. The skills do not file applications, do not sign inventor declarations, do not submit Information Disclosure Statements, do not contact the USPTO and do not appear before any tribunal. Signing a declaration under penalty of perjury, the duty-of-candor and IDS materiality call, recording an assignment, paying the PTO and anything requiring a registered practitioner remain reserved to a human, and the README says the skills refuse those acts without separate explicit authorization.

The failure modes named in the README are the ones to plan around. Models hallucinate cites, miss art and invent confidence. Fee amounts, form names, PTO mechanics and 101 case law move, so the examination and workaround skills are instructed to refresh statutes, rules, MPEP pages and examination memoranda on every run, and the README tells you to verify those sources yourself. It also names the consequences plainly: a bad claim set, a missed statutory bar, a false entity-status claim or an omitted material reference can destroy rights. This is the wrong tool if you cannot check primary sources, or if the stakes are high enough that a missed deadline is something you would sue over. The README's own answer for that case is to hire a registered practitioner and use the repository to make that person cheaper.

How this differs from a patent drafting tool or a legal chatbot

The README contrasts the repository with a chatbot wrapper and with a talk-to-a-lawyer landing page, and the structural difference is real: the product is files, not a service endpoint. A drafting tool typically takes a disclosure and returns prose. Here, patent-audit returns ranked defects and a parsed.json, patent-examine returns a version-locked candidate only after a simulated Office Action, and patent-workaround returns an adversarial attack on the claims. The gate is procedural rather than generative.

The second difference is the counting. Deterministic Python handles the mechanical checks, and the README says the scripts exist precisely because models cannot count or do date arithmetic. A general-purpose legal chatbot has no parser building a claim dependency graph and no 64-item claims part that agents are forbidden to skip. The trade-off is that the pipeline is opinionated: three skills run in a fixed order, each with its own SKILL.md contract, and the README states that skipping a stage is not a substitute for running it. If you want a single prompt that drafts an application, this is not that, and the README does not pretend otherwise.

Maintenance, licence and the cost of keeping the files current

The repository is not archived and the last push was on 2026-08-12. There are no releases retrieved, so there is no versioned artifact to pin; you track the default branch. The README argues the maintenance model is the point: if a better model ships, the same files get cheaper and sharper, and if the PTO changes a form, you edit a markdown file instead of waiting for a CLE. That is a genuine advantage over a hosted tool whose behaviour you cannot edit, but it moves the upkeep to you. The examination and workaround skills are instructed to re-fetch statutes, rules, MPEP pages and examination memoranda on every run, which means your results depend on sources outside the repository being reachable and current at run time.

The licence is GPL-3.0, stated in the README and present as LICENSE in the top-level layout. The README notes that the GPL-3.0 license already disclaims warranty. Copyleft obligations attach if you redistribute the skills or a derivative, which matters if you were planning to fold them into a proprietary internal product; whether your specific use triggers those obligations is a question for your own counsel, and nothing here is legal advice. There is no homepage, so the repository itself is the distribution channel.

Editorial conclusion

Adopt it if you are an inventor, founder or in-house engineer who can check statutes, 37 CFR and MPEP text yourself and wants a pre-filing audit, a mock prosecution loop and a design-around attack before paying anyone. Do not adopt it as a filing service: the README states the skills do not file, do not sign declarations, do not submit IDSs and do not contact the USPTO, and the acts that need a registered practitioner stay reserved to a human. Verify first that the three skill directories install where your agent looks for SKILL.md, that the deterministic scripts run on your Python, and that every statute, rule and MPEP page the agent re-fetches on each run is current before you rely on a result.

Frequently asked questions

Official sources

1. gfodor/legal-skills on GitHub
2. Issues
3. License: GPL-3.0
4. README