I'm a patent attorney (USPTO Reg. 46,376). I built the software I wish every inventor who couldn't afford me had had.
It won't be your lawyer. You file it yourself. But it does the prior-art search and the first draft, and it shows you every step. Here's what it does, and what it deliberately doesn't.
The problem isn't that inventors are lazy. It's the price. A non-provisional from a firm runs $9K to $17K. Most solo inventors either don't file, or file something so narrow it's worthless. Filing it yourself is legal. Doing it well, alone, is the hard part.
What Abigail does:
Searches prior art (USPTO, EPO, Google Patents) and shows you the references.
Drafts a spec, independent and dependent claims, and drawing descriptions.
Shows, for every claim, which of your own words support it and why it's worded that way.
You read it, you decide, you file via Patent Center.
What it deliberately does NOT do:
It is not your lawyer and gives no legal advice.
It won't file for you.
It is not for litigation or high-stakes M&A work.
It will tell you when your idea probably isn't patentable, instead of taking your money. That last one is the whole point.
The fear I hear most: "if I type my invention into an AI, do I lose it?" Real concern. With many tools your input can become training data, and public disclosure can cost you rights. Abigail does not train on your inputs and does not retain them. That was non-negotiable for me.
On accuracy: any vendor claiming "hallucination-free" AI is overselling. A Stanford study caught the big legal-AI tools getting it wrong 17 to 33% of the time. So Abigail doesn't ask you to trust it. It shows the source behind every claim so you can check it. That's the Glass Box.
Cost, plainly: $25 in free credits when you sign in, and the price is on the screen, no sales call, versus $9K to $17K at a firm. Cheap because it's software, not billable hours.
If you've been sitting on something you couldn't afford to file: https://t.co/GCu22HED4W. It'll tell you honestly if it isn't worth filing.
- Roger
One rule in our benchmark scares the vendors more than all the others. It should.
If a tool invents a citation that would go into a filing, it doesn't lose points. It fails. Zero. A perfect score everywhere else can't buy it back.
We named it after Therasense, 649 F.3d 1276. Every prosecutor reading this just winced, because we know what a fabricated reference does to a patent, and to the person who signed it.
Most AI benchmarks let a tool average its sins. A little hallucination here, great speed there, round it up to a 92. Ours doesn't round. Cross the line on candor and the tool scores zero, however fast or fluent the rest of the run was.
That's what a benchmark looks like when the people writing it have to sign the response. Not the people selling the tool.
https://t.co/1wDNAcMBZS
Patent eligibility under Section 101 moved more in the past year than in the decade before it. The framework changed. Most AI patent tools did not.
They draft to a generic standard with one general model doing everything. The hardest question in software and AI prosecution, whether the claim survives 101, gets the same shallow pass as formatting the abstract.
Abigail does not run one model. It runs eleven, each a specialist: 101 eligibility, 102 novelty, 103 obviousness, 112 support, claim drafting, prior-art mapping, amendment strategy, examiner analytics, and more. The eligibility specialist is tuned to the practical-application test the Office actually applies, not a generic idea of what a patent should look like.
You see all eleven. Where they agree you have a strong argument. Where they diverge you have a decision to make, with the reasoning in front of you.
I run this on real cases because I prosecute real cases. It was built by a practitioner around the rejections that actually come back.
https://t.co/vrUY6PkJiL
The founders of the biggest AI patent tools have never prosecuted a patent. Here is how they cover for it. They rent the credential.
One was founded by a venture capitalist and a crypto engineer, then announced an advisory board of more than fifteen patent people, one press release at a time. A former USPTO Director. Big Law IP chairs. Every one an advisor. None building the product. None on the paper when you sign.
Another, founded by two AI engineers, markets itself as built by patent professionals for patent professionals. Read that carefully. It does not mean the founders. It points to an outside advisory board.
A third, founded by three machine-learning PhDs, gives early hires the title Founding Member of Legal. They joined after the engineers founded the company.
This is the tell. When the people who build the tool have never done the work, they bolt the work onto a letterhead. An advisory board reviews slides. It does not draft your claim amendment, it does not face your examiner, and it is not on the paper when you sign.
I did not add a patent attorney to a board. I am the patent attorney. I built the tool, and I use it on my own cases.
Ask any vendor one question: who on your founding team has ever prosecuted a patent. Then watch the room.
https://t.co/vrUY6PkJiL
@MarioNawfal Political science isn’t a science. Just listen to this guy. Escalation traps, dueling frameworks, strike prices, outside options. I wonder what important sounding phrase he’ll reach for next to dress up a guess as a finding. 🙄
Most founders think filing the patent application is the finish line. It's mile one. Here's the road, so nothing on it surprises you.
You file. Then you wait. The examiner queue runs roughly 20 months before anyone picks up your case. You're patent pending that whole time, which carries real weight with investors and on your site.
Then the first office action shows up, and it's almost always a rejection. Nearly 9 in 10 applications get rejected at least once. Read it as the examiner telling you exactly what stands between you and a grant.
You get 3 months to respond, extendable to 6. A good response amends the claims, argues the cited art, or both. Some cases clear in one round. Budget for two.
Then allowance, the issue fee, and a patent number. Idea to grant runs 2 to 3 years for most software and AI cases.
The applicants who do well aren't the fastest filers. They're the ones who walked in knowing the whole road.
https://t.co/GCu22HED4W
An office action is not a rejection of you. It is the examiner's opening position.
The examiner cites specific prior art against specific claim limitations. Read the actual references, not the form paragraphs. Often the mapping has a gap, and that gap is your argument.
You can respond yourself. The deadline is real, the fee is small, the writing is the hard part.
https://t.co/GCu22HFaUu. Not legal advice.
A provisional patent application costs $65 to file with the USPTO if you qualify as a micro entity. Not $9,000.
The $9,000 is what a firm charges to write it. The filing date is what you are actually buying, and in the US the date you file beats the date you invented.
What stops most solo inventors is not the fee. It is the prior art search and the claims. That is the part https://t.co/GCu22HED4W helps you do yourself.
Not legal advice.
Your first office action arrived and it's a rejection. Good. Now you know what the examiner actually thinks.
The basic anatomy. A 102 says one reference already shows your invention. A 103 says a combination of references makes it obvious. A 101 says the claim is too abstract, the big one in software and AI. A 112 says the application doesn't describe or claim the invention clearly enough.
Three moves, in order.
First, read the cited art. Not the form paragraphs, the actual references, against your actual claim language. Half the time the examiner's mapping has a gap, and that gap is your argument.
Second, call the examiner. An interview takes 30 minutes and it's the highest-leverage half hour in prosecution. You learn what the examiner really objects to and which amendment would clear it.
Third, choose: amend or argue. Amending narrows the claim and trades scope for allowance. Arguing preserves scope but has to engage the references head on. The classic mistake is arguing features that aren't in the claims. Examiners ignore that, correctly.
Sitting on a rejection right now? https://t.co/GCu22HED4W
Pliny jailbroke Fable 5 using Parseltongue.
Yes, y0u r3aD th@t r1ght.
Parseltongue.
Hours later, the US government banned the model.
At 5:21pm ET, Commerce Secretary Howard Lutnick sent Anthropic an export control directive. Suspend Fable 5 and Mythos 5 for every foreign national, inside the US or out, including Anthropic's own foreign born staff. You can't comply with that halfway. So Anthropic shut both models off for everyone (https://t.co/P3LilDkgIz).
Three days. That's how long the most powerful AI in America lasted in public.
What set it off? Per reporting, a competitor told Commerce it could break Mythos. Anthropic says the "jailbreak" was just asking the model to read a codebase and fix software flaws. Probably using the same techniques publisheed by Pliny.
However, Anthropic says the same trick works on OpenAI's GPT-5.5, and if this becomes the standard, it "would essentially halt all new model deployments for all frontier model providers."
This is different from the TikTok ban. US gov used export controls, the Bureau of Industry and Security, the same rule from the 1990s software code wars. Under the export rule, 15 CFR 734.13, letting a foreign national in the US touch controlled code is an export to their home country.
In Bernstein v. United States, a court held software source code is protected speech, and that export licensing of it was an unconstitutional prior restraint. Junger v. Daley agreed. Then TikTok v. Garland let a ban stand, but only on a narrow, specific national security showing.
So two worrisome implications: (1) The government can switch off the smartest tool in the country in an afternoon, on a rival's say so! (2) basis is forty years of unsettled fights over whether code is speech.
I published a series this week on Fable's safeguards, lack of zero data retention and effect on client privilege, and the jailbreak iself.
Download whitepaper here: https://t.co/l1UqXgfeFw
What happens when the next directive skips the press release? https://t.co/6tBHnJ5yiR
Yesteday, I posted that a researcher who calls himself Pliny jailbroke Fable 5 using Parseltongue.
Yes, y0u r3aD th@t r1ght.
Parseltongue.
Hours later, the US government banned the model.
At 5:21pm ET, Commerce Secretary Howard Lutnick sent Anthropic an export control directive. Suspend Fable 5 and Mythos 5 for every foreign national, inside the US or out, including Anthropic's own foreign born staff. You can't comply with that halfway. So Anthropic shut both models off for everyone (https://t.co/P3LilDkgIz).
Three days. That's how long the most powerful AI in America lasted in public.
What set it off? Per reporting, a competitor told Commerce it could break Mythos. Anthropic says the "jailbreak" was just asking the model to read a codebase and fix software flaws. Probably using the same techniques publisheed by Pliny.
However, Anthropic says the same trick works on OpenAI's GPT-5.5, and if this becomes the standard, it "would essentially halt all new model deployments for all frontier model providers."
This is different from the TikTok ban. US gov used export controls, the Bureau of Industry and Security, the same rule from the 1990s software code wars. Under the export rule, 15 CFR 734.13, letting a foreign national in the US touch controlled code is an export to their home country.
In Bernstein v. United States, a court held software source code is protected speech, and that export licensing of it was an unconstitutional prior restraint. Junger v. Daley agreed. Then TikTok v. Garland let a ban stand, but only on a narrow, specific national security showing.
So two worrisome implications: (1) The government can switch off the smartest tool in the country in an afternoon, on a rival's say so! (2) basis is forty years of unsettled fights over whether code is speech.
I published a series this week on Fable's safeguards, lack of zero data retention and effect on client privilege, and the jailbreak iself.
Download whitepaper here: https://t.co/l1UqXgfeFw
What happens when the next directive skips the press release? https://t.co/6tBHnJ5yiR
Patents rarely die on the merits. They die on calendars.
Three clocks, and most inventors only know about one.
Clock one starts before you file. Demo at a conference, publish a paper, or offer the product for sale, and a 1-year clock starts on your US rights. Miss it and your own disclosure becomes prior art against you. Most foreign rights are gone the day you disclose, no grace period at all. If a launch is coming, file first. Even a provisional buys you 12 months.
Clock two runs during prosecution. An office action gives you 3 months to respond, extendable to 6 with escalating fees. Miss month 6 and the application goes abandoned. Getting it back takes a petition, a fee, and a sworn statement that the delay was unintentional.
Clock three runs after you win. Maintenance fees come due at 3.5, 7.5, and 11.5 years. Skip one and the patent lapses while your competitors are still reading it.
The fix is boring and it works: calendar every clock the day it starts, with a reminder a month out.
Not sure which clocks are already running on your invention? https://t.co/GCu22HED4W