So long, dumb tools
For most of software’s history, the tool awaited the human expert. PowerPoint didn’t design decks, Salesforce didn’t close deals, Photoshop didn’t choose color palettes. Mastery lived in our heads; software merely stored or executed our decisions.
Large language models are changing that dynamic. A well-crafted prompt distills an expert’s playbook into a reusable artifact. When that prompt is embedded inside the product, a new user isn’t just using software—they’re borrowing expertise on demand.
Developers felt this first. GitHub Copilot ships with fragments of millions of senior engineer keystrokes and now has roughly a 30% suggestion-acceptance rate inside enterprises—and that percentage climbs as trust compounds (GitHub). But dev tools were only the start.
Prompts are the new programs
A prompt is, functionally, compiled judgment. In dev tools, there’s an ongoing stream of system prompt leaks from tools like Cursor, Windsurf, and v0 (X). Why do these leaks provoke such a strong reaction? Because they underscore the fact that so much of the value of these products is the prompt—the encoded best practices and guidance—used to steer model outputs.
Claire Vo’s ChatPRD is one non-developer example, condensing a decade of CPO-level product judgment into a chat assistant that drafts and critiques product-requirements docs. Demand is real: ChatPRD’s weekly revenue recently exceeded its entire June 2024 haul, and Claire soon after announced that she was leaving her role as CPTO at LaunchDarkly to go full-time on the project (X).
The UI that teaches while it works
The most radical software won’t simply show users best practices; it will steer them along the optimal path through a combination of recommendations and assisted execution.
In a hybrid state, UI elements—defaults, required fields, feedback loops—can guide users toward choosing the path a top-tier operator would take. In a fully agentic world, expert workflows can execute in the background, conducting autonomous A-B tests, allocating spend across channels, or generating content for a specific prospect. SaaS 2.0 tools like MixMax and Iterable already use AI to send emails at recommended times based on historical data; agentic SaaS takes this one step further.
Other examples:
A toggle called “Standardize Pricing” applies the discount guardrails your best rev-ops leader would set.
A default “5-day nurture” email cadence reflects historical data on lead scoring.
A linting rule warns the developer when an endpoint strays from a scalable API design pattern.
Where does this lead? Toward democratization. Teams of five ship what once took teams of fifty, because every seat comes bundled with an expert agent. With the falling cost of user research thanks to AI-assisted tools like Listen Labs, companies can scale their expertise collection efforts and ingest more data than ever around how their products should feel and work. Curation of expertise may become its own role—an evolution of the PM?—as gathering, auditing, and updating best practices grows even more important.
Evals evals evals
What does “good” look like? That’s the question builders are perpetually trying to answer in order to embed specific domain expertise into models. The importance of eval data sets will only continue to grow, as will the importance of data flywheels created by user telemetry. In the age of fierce competition in AI-powered SaaS, you’re only as good as your rate of improvement.
The takeaway
If the 2010s were about “software eating the world,” the 2020s are about software absorbing expertise. As that trend accelerates, the real differentiator will be how quickly a product can incorporate new data, update its prompts, and arm users with the judgment they need to achieve their goals.