Continuation Drafter
Continuation claims, drafted on your own machine
A specification, the filed parent claims and, if you have it, the parent's file history go in. Drafted continuation claims come out, with their defects already flagged, so you start from a marked-up draft instead of a blank page.
Not ready to install anything? Start with the free course: continuation practice in the order a family is built, over authorities you can check for yourself.
Free. Runs locally through Ollama, or through your own provider key. Nothing to activate. Every run ends by reporting what it cost.
Prefer a window to a prompt? One command opens a local interface in your browser, with the same drafting, critique and scoring.
The claims above are written to show the shape, not produced by the tool. A real run is printed unedited here, including every defect the tool flagged in its own output.
Where your specification goes
An unpublished application is confidential client material, and where it ends up is a question worth settling before any question about how good the drafting is. Here is the whole answer, including the part that is a caveat.
On the local path, inference runs on your hardware and the specification is not transmitted anywhere. That property belongs to the local path, not to the tool as a whole: point it at a remote provider and the specification reaches that provider, which suits published material and is a decision you make per matter for anything unpublished.
Local inference is not unique to this tool. What is unusual is the shape: on a Mac, an app you download and double-click today, rather than a deployment behind a sales cycle and a security review. Command-line binaries for macOS, Linux and Windows are published beside it.
Numbers you can check, not logos
The two closest funded competitors prove themselves with customer logos and testimonials. This one has none, so it proves itself the only way that is honest for a new tool: by publishing the method and letting you disagree with it.
Drafting quality
Wall-clock, a separate measurement
The timed runs are a subset: 15 of the 51 models, all remote, so local runs on your own hardware will be slower. This measures how long the tool takes, not how much time it saves you: that is a different claim, we have not measured it, and so we do not make it. The method, and what it does not tell you.
Against real prosecution records
What real practitioners did, measured from public USPTO data, not a recommendation about any application. The three measurements, their method, and a finding against the tool.
The authorities, quoted and free to read
The same argument as the numbers above, in a register that does not depend on trusting us at all: 48 authorities on continuation practice, each quoted from its primary source, reporting what the authority held and how it bears on drafting. No account, no email, nothing to install.
Read the corpus, or take it in the order a family is actually built with the free course, which cites into these records rather than replacing them.
Written for licensed practitioners and those working under their supervision. These pages report what an authority held; they are not advice about any matter, and they do not tell you what to file.
What it actually does
Four things, and it is honest about which of them is solid.
Drafts
Reads the specification and the parent claims, then drafts continuation claims aimed at what the disclosure supports beyond the parent's scope.
Reads the file history
Attach the parent's Patent Center "download all documents" PDF. It picks out the Office Actions, the responses and the Notice of Allowance, reads scanned pages on your own machine, and reports each rejection and the examiner's reasons with the quotation each came from. Every limitation of the allowed claims is marked as filed with the application or added in a round of prosecution. What that means for the continuation is yours to judge.
On real Patent Center file histories, a 9-billion-parameter model on a laptop found every rejection, and no model named a reference the paper does not contain. How that was measured.
Critiques
Flags per-claim 112(a) and 112(b) defects, each with a proposed fix. In our own testing this was the most useful thing it produced.
Scores
Rates drafting quality against the specification you supply: element and combination support, definiteness, and how a claim differs from those already filed. Read the spread rather than the rank: a single AI judge cannot separate near-equal drafts, and we say so on the benchmark page.
51 models, ten real patent families, one table
We ran the same ten parent and continuation pairs through every model we could host, and the interesting gap was not the one we expected.
| Your machine | Model | Score | Finished |
|---|---|---|---|
| A server with 512 GB of memory | glm-5.2 | 88.4 | 10 of 10 |
| A workstation with 64 GB | gemma4:31b | 81.3 | 9 of 10 |
| A desktop or laptop with 32 GB | qwen3.8:27b | 85.5 | 8 of 9 tried |
| A laptop with 16 GB | gemma4:12b | 67.3 | 9 of 10 |
The gap between memory tiers is larger than the gap between open and closed weights. A firm running the top row sits about 5.0 points behind the strongest closed-weight model measured, with nothing leaving the building. How this was measured, and what it does not tell you, or which model to choose in plain terms.
And then it is back in Word, where you actually work
The drafted claims are text. One command puts them in a document you can edit, which since 2024 is the format the USPTO expects for the description, claims and abstract: filing those in any other format carries a surcharge.
Your claim numbering is preserved exactly as drafted and never renumbered, since a continuation's claims are commonly numbered on from the parent. LaTeX works too. The exported file carries the maturity notice inside it, because an exported document gets read by people who never saw the terminal that made it: it is a working draft for you to edit, not a submission.