Patent Office Actions
Reply drafting, measured against filed replies

A first draft of the reply, built on the record

Patent Practitioner Tools drafts each item of an Office Action as parts you adopt, edit or discard, from the examiner's own reasoning, the examiner's citations and the references' own words, with every quotation checked against its source. We gave it 10 real Office Actions and compared its drafts with the replies practitioners filed.

Free to download and use, for macOS, Windows and Linux. Draft on a model on your own computer, so the file history and your draft stay on it, or on your own provider key. No account.

What the drafts raised

Each draft was scored for which of the 87 points in the filed replies it makes, and each drafted part for whether a practitioner could use it after light editing.

Reply drafting, measured against filed replies · 10 Office Actions

20 of 87 filed points raised by claude-opus-5.5 15 of 87 by gemma4:31b, running locally 83% and 98% of claude-opus-5.5's parts usable as a draft after light editing, by each judge

Points counted where both judges agree. Every drafter, both judges.

Why start the reply here

The reply stays yours to write, sign and answer for. What changes is where you start.

Built on the record you already have

It reads the Office Action and the papers you add, or the whole file history you download from Patent Center. Each item is drafted from the examiner's statement and reasoning, the examiner's citations, and, for each reference you have, its own words nearest your claim's language. Nothing to retype.

Every quotation checked

Before a part is shown, its quotations are looked for in the record, the specification and the references in hand, as they were quoted: a quotation shortened with an ellipsis is found part by part, in order, and one across a page's header and footer is found across it. A part quoting words found in none of them is withheld and says so, and a sentence stating an outcome is removed.

Yours, one unit at a time

Text marked From the record is written from the prosecution record without an AI model; Suggested text is AI-generated; Review marks text a check flagged or whose review scope you have not confirmed. All of it requires practitioner verification. Each unit goes into your reply with Add to draft, or is edited or dismissed; an added unit can be taken out again, and each keeps its history. Added text goes under the item's heading, the checks run over what you wrote, and the reply exports to .docx with a sheet of what the checks found, to read and delete before filing.

On your own machine, if you want

Draft with a model on your own computer, so the file history and your draft stay on it, or with your own key at a provider you choose. The same model reads the record and drafts the reply; what was measured helps you pick one for your memory.

Each item, the approach you choose

Every rejection, objection and requirement is its own workspace under Address the Rejections and Objections, with the list of items beside the one you have open on a wide screen.

An approach for each item

Chosen before anything is drafted, and offered by the kind of item: arguments, amendments, both or canceling the claims for a rejection; a terminal disclaimer as well for nonstatutory double patenting; for statutory double patenting no terminal disclaimer, with MPEP 804.02 quoted beside the list; the correction or an argument for an objection; an election, with or without traverse, for a restriction. Every list ends with Other, in your own words, which the drafter is given as you wrote them.

What a statement of absence rests on

A statement that a reference does not describe something cannot be quoted, so it goes in only once you say what you reviewed with Confirm review scope: the passages the examiner cited, the passages shown, or the whole of each reference. Beside it are the passages it rests on and a note that the rest was not read here. It is a record of what you reviewed, not a finding about the reference.

The reply as a document

Read the reply as a Document, section headings in bold, each with what the last check found for it, and what changed since your last export. The check leads with its counts: the items with matching draft text detected, the quotations found in the sources in hand, and the amended claims whose added words were all located in the specification.

Notes that stay out of the export

A note goes on one section of the reply or on the whole of it, and is resolved rather than deleted. Open notes are counted under the export buttons, on their section and in Next steps, and none is ever written into the exported reply. They stay on your computer with the matter.

The round, drafter by drafter

10 prior-art rejections, and 87 points in the replies practitioners filed to them: 56 arguments, 19 amendments and 12 about dependent claims. Points show judge 1 (gpt-5.5), judge 2 (gemini-3.1-pro-preview) and, in bold, both; usable parts show each judge. For scale, a draft that only restates the claim and quotes the examiner raises 1.

Reply drafting, measured against filed replies: 10 Office Actions, run 2026-10-04 on the drafter released as 0.7.0
DrafterRunsFiled points raised, of 87Filed arguments raised, of 56Parts usable after light editingParts withheld by the quotation check
claude-opus-5.5at a provider27, 21, 201083%, 98%0
gpt-5.6-solat a provider22, 15, 14586%, 95%0
gemma4:31blocally20, 15, 15746%, 76%1
qwen3.8:27blocally14, 15, 11667%, 96%5
qwen3:8blocally13, 13, 9343%, 63%11

Arguments are where drafters differ

Every drafter starts from the same statement about the dependent claims, read from the claims with no model, so the arguments column is where they separate: 10 of the 56 filed arguments for claude-opus-5.5, 7 for gemma4:31b.

On your own machine, as many filed points as a frontier model

gemma4:31b, running locally, raised 15 of the filed points to 14 for gpt-5.6-sol, and 7 filed arguments to 5, with the file history and the draft staying on the computer. Its parts take more editing: 46% and 76% were usable as drafts.

The frontier drafters

claude-opus-5.5 raised the most: 20 of the filed points and 10 filed arguments, with 83% and 98% of its parts usable as drafts. gpt-5.6-sol was about as usable (86% and 95%) and raised 14.

Every quotation checked, here too

The quotation check ran before the judges saw anything, as it does in the app, so a part that misquotes its source never reaches the draft. It withheld no part from the frontier drafters, 1 from gemma4:31b, 5 from qwen3.8:27b, 11 from qwen3:8b.

How it was measured

  • Real filed replies. For each of 10 public prosecution families, the record was cut just before the first filed reply to a prior-art rejection, so every drafter saw what the practitioner saw, and drafted a reply to that rejection.
  • Scored against what was filed. A judge model listed the points the filed reply made, and each draft is credited with the ones it raises. A different good argument earns nothing, so the count measures agreement with the reply that was filed.
  • Two model judges, both shown. gpt-5.5 and gemini-3.1-pro-preview, at temperature 0, scored every draft and judged each part as a draft, never whether it would persuade. gpt-5.5 shares a vendor with gpt-5.6-sol, so judge 2's figure is the cleaner one for it.
  • As in the app. The quotation check ran first, and every drafter used the drafter the program ships, one call per item.

Ten Office Actions show large differences clearly. The round ran on 2026-10-04 on the drafter released as 0.7.0; since then the drafter has gained drafting each item from the examiner's own reasoning (0.7.1), the heading, opening statement and closing request a filed reply carries (0.7.3), and drafting each item on its own with the arguments, amendments or both chosen for it (0.9.0), and amendments written into the claim they name, amendments to the specification, its title and its abstract, and arguments on a section 103 rejection that address the examiner's combination as a whole (0.10.1), and arguments asked to rest on what the references describe rather than only to ask the examiner to explain, with an item drafted again from the whole of a reference when you ask (0.11.0), and arguments drafted against your amended claims once the reply amends them (0.12.0). The figures above are the 0.7.0 drafter's; those later additions are not in them. This measures drafting a reply: /models measures drafting continuation claims, and /prosecution-evidence reading an Office Action.

The figures on this page are published under CC BY 4.0: reuse them, with credit to Obviously Not.

Try it on an action you know

Add an Office Action you have answered before, or a public Patent Center file history you know well, ask for a draft of the reply, and read each part against what you would file. Nothing to sign up for, and your documents and drafts stay on your computer if you draft locally.