Anthropic's Formal Verification Post Shows the Growth Play: Drown the Reader in Specific Numbers, and Skip the Adjectives Entirely.

The Play

Anthropic's formal verification post could have opened with "AI is transforming software security." Almost every AI lab's blog post does some version of that sentence.

Instead it opens with specific, checkable facts: 239 entry points proved safe, across 123 different libraries, generating 180 million lines of proof code, with individual proofs running up to 600,000 lines. It tells you the binaries verified were up to 18,000 RISC-V instructions long, and that proving them required staying inside a 64 MB input limit.

None of that is marketing copy. It reads like a lab notebook. That's the point.

Why specificity outperforms hype

A reader cannot verify "groundbreaking." They can verify "239 entry points across 123 OSS-Fuzz libraries" by clicking through to OSS-Fuzz and checking the libraries themselves. The moment a claim becomes checkable, skepticism drops — and sharing goes up, because the person resharing it isn't vouching for an opinion, they're relaying a fact.

The post's strongest line does this twice in one sentence: Claude found bugs that "eluded detection for years, despite trillions of test runs." That's not an adjective doing the work — it's a direct, almost embarrassing admission about how long fuzzing missed something, which makes the claim about AI's capability land harder than any superlative could.

Anthropic also wasn't afraid to publish unflattering-looking numbers next to the impressive ones: it took 20 person-years to formally verify seL4 by hand, versus 2 person-years to build it in the first place. That specific, slightly damning cost comparison is exactly what makes the "AI collapses that ratio" claim credible — you don't trust a company that only ever shows you the flattering number.

How to steal this

Before you publish your next technical post, product update, or case study, go through every adjective and ask: what's the actual number behind this word? If you can't answer that, either find the number or cut the claim.

• "Fast" → what's the actual latency, in milliseconds, compared to what baseline?
• "A lot of users" → what's the actual count, and over what time window?
• "Big cost savings" → what's the actual dollar figure or percentage, sourced from where?

If the honest answer makes the claim look smaller, say the smaller true thing. Readers forgive modest. They don't forgive vague.

Bottom line

Anthropic's post didn't need adjectives because the numbers were doing all the persuading. That's a content strategy any founder can copy for free: before you ship the next post, strip every intensity word and replace it with the receipt.

Source: https://www.anthropic.com/research/formal-verification