The humble Product Requirements Document (PRD) has been the backbone of software development for decades. These lengthy documents outlined features, user stories, and technical specifications in painstaking detail before a single line of code was written. But according to Aparna Chennapragada, Chief Product Officer at Microsoft, that era is rapidly coming to an end.
"I call it the prompt sets are the new PRDs," Chennapragada said in a recent interview. "I really insist on folks saying if you're building new projects, new features of course come with prototypes and prompt sets."
This isn't just semantic wordplay or another Silicon Valley buzzword pivot. It represents a fundamental shift in how product teams approach development in an era where Natural Language Experience (NLX) is becoming the dominant interface paradigm. As Chennapragada puts it: "NLX is the new UX."
The death of "memos before demos"
The traditional product development cycle has been turned on its head. Where teams once spent months crafting detailed specifications, Chennapragada advocates for immediate prototyping: "In this day and age, if you're not prototyping and building to see what you want to build, I think you're doing it wrong."
The philosophy she champions—"demos before memos"—leverages AI's ability to rapidly transform ideas into working prototypes. But these aren't traditional clickable mockups. In the age of conversational interfaces and AI agents, the most useful product spec isn't a PDF—it's an executable set of prompts that can generate live, interactive demonstrations of intended behavior.
"It's a much more high bandwidth way of communication," Chennapragada explained. Rather than hoping developers correctly interpret written requirements, teams can now show exactly what they envision through working conversations with AI systems.
What exactly is a "prompt set"?
A prompt set functions as an executable specification that bundles several key components:
- Canonical prompts and roles (system/user interactions)
- Variations for edge cases and user personas
- Evaluation prompts with acceptance criteria (both automatic and human-rated)
- Guardrails defining what the model must and mustn't do
- Context contracts specifying what data and tools are in scope
- UI expectations for NLX (such as showing plans before acting, progress indicators, follow-up suggestions)
If a traditional PRD describes what to build, a prompt set demonstrates how it behaves—today, with current models, with testable outcomes.
This approach addresses a fundamental shift in how users interact with software. As Chennapragada notes, "conversations also have grammars, they have structures, they have UI elements, they're invisible." Product teams must now design these conversational structures just as deliberately as they once designed buttons and menus.
The new design language of conversation
Natural language interfaces introduce entirely new UI elements that require explicit design decisions:
Plans as UI elements: When users give high-level goals to AI agents, the system should respond with editable plans rather than immediately taking action. This gives users visibility and control over what's about to happen.
Progress narratives: How much should the system reveal about its thinking process? Too verbose feels like watching server logs; too terse leaves users uncertain whether the system is working correctly.
Follow-up suggestions: After completing a task, what are the "next obvious steps" the system should proactively suggest? This guides users toward successful workflows without being intrusive.
Confidence indicators: When should the system cite sources, express uncertainty, or ask for permission before proceeding?
These elements emerge from recent design guidance for agent UX systems, formalizing what was once ad-hoc into designable, specifiable patterns.
Anatomy of a Prompt-PRD
A comprehensive prompt-based specification includes several key sections that traditional PRDs lack:
Core Problem & Flows
- Problem statement & jobs-to-be-done in plain language
- Canonical flows with specific prompts and expected UI behaviors
- Kickoff prompt defining role, tone, and guardrails
- Planning behavior specifying how agents should propose editable plans
- Progress behavior defining when and how to narrate work
- Follow-up patterns outlining contextual next steps to suggest
Technical Contracts
- Context contract defining available data sources, APIs, and permission boundaries
- Evaluation suite with test prompts and expected outcomes
- Safety policies for handling unsafe requests, jailbreak attempts, and prompt injection
- Versioning information including model versions, parameters, and data indices
A Minimal Example
Consider a "One-page PRD generator" feature:
- Kickoff prompt: "You are a senior PM assistant. Ask 3 scoping questions first, then output sections: Overview, Goals, Non-Goals, NLX flows, Metrics."
- Plan behavior: Always propose a 3-step plan before writing
- Progress behavior: Show one status message if processing takes >5 seconds
- Follow-ups: Offer to (a) generate prompt set, (b) synthesize exec summary, (c) create evaluation tests
- Context: Can read linked docs and issue trackers; no PII retrieval
- Safety: Refuse legal/medical claims; strip injected instructions
How this changes the development process
This shift creates what Chennapragada describes as an "uneven cadence" in product development. While the time to create initial demos has dramatically shortened, the path to full deployment may actually take longer due to higher quality standards.
"The time to first demo is much shorter, but the time to a full deployment is going to take longer," she noted. Teams can rapidly explore multiple concepts but must be more selective about which ideas warrant full development.
The abundance of easily-created prototypes raises the bar for what constitutes compelling products. "There's going to be a supply of ideas, a massive increase in supply of ideas in prototypes which is great. It raises the floor, but it raises the ceiling as well."
New Development Practices
Design reviews add NLX checkpoints: Teams must now decide how much planning to show, what level of verbosity is appropriate, and which follow-ups are helpful versus spammy. These become explicit design decisions, not model quirks.
Shipping includes evaluations: Continuous integration can run prompt tests against nightly builds—especially important as model versions change. Teams need systematic ways to test conversational behaviors.
Roles blur at the edges: The traditional boundaries between product managers, designers, and engineers are shifting as AI tools democratize prototype creation. However, this makes editorial judgment more crucial, not less.
The Microsoft "Frontier" approach
Microsoft has institutionalized this philosophy through what Chennapragada calls the "Frontier program"—putting cutting-edge AI tools in the hands of early adopters to help them experience working "one year in the future."
The company has even established a fake company where teams can experiment with advanced research agents and cutting-edge tools without disrupting critical enterprise operations. This addresses a key challenge in enterprise AI adoption: the tension between rapid technological advancement and careful change management.
Security and safety considerations
The shift to conversational interfaces introduces new security challenges. Visible "thinking" processes can leak sensitive context, and chain-of-thought reasoning creates larger attack surfaces for prompt injection and jailbreaking attempts.
Security teams have published specific advisories about reasoning models like DeepSeek-R1. When prompt specifications require "show your work" behavior, teams must scope it carefully, prefer summaries over raw reasoning traces, and test against hostile prompts.
There's also an ongoing debate about whether models are truly reasoning or sophisticated pattern-matching—and what this means for reliability. Prompt-PRDs should explicitly define when humans must remain in the loop.
The skeptical view
Not everyone agrees that traditional PRDs are obsolete. Critics argue that comprehensive documentation remains essential for complex enterprise software, regulatory compliance, and long-term maintainability. The risk of "Frankenstein products" that Chennapragada acknowledges suggests that rapid prototyping without proper planning can lead to incoherent experiences.
Questions remain about whether this approach scales beyond small, experimental teams. While a three-person team with advanced AI tools can move quickly, translating these practices to larger organizations with complex stakeholder management needs presents challenges.
Implementation checklist
For teams ready to adopt prompt-based specifications, here's a practical starting framework:
- Problem & jobs-to-be-done
- Kickoff prompt (role, tone, guardrails)
- Plan pattern (editable plans before action)
- Progress pattern (when/how to narrate work)
- Follow-ups pattern (contextual next steps)
- Citations/confidence rules
- Context contract (data, tools, permissions)
- Evaluation suite (automated and human testing)
- Safety policy (refusals, injection handling)
- Versioning (model, parameters, data indices)
- Rollout strategy (frontier users first, logging for model drift)
Looking ahead
As AI capabilities continue advancing, the gap between idea and implementation will likely continue shrinking. Chennapragada's vision of prompt sets replacing PRDs may seem radical today, but it reflects broader trends toward more immediate, experiential product development.
The success of this approach depends on teams' ability to maintain product quality and coherence while embracing rapid experimentation. For organizations willing to adapt their processes, the promise is compelling: faster innovation cycles, better communication between team members, and products that more closely match their original vision.
Mainstream platforms are already incorporating NLX into build flows, and design organizations are publishing agent-UX principles that product teams can adopt wholesale. The infrastructure is catching up to the rhetoric—which is exactly why prompt-PRDs are becoming useful: they pin conversational concepts to working behavior.
Whether prompt sets truly replace PRDs or simply complement them, one thing is clear: conversation has become a designed surface that requires the same deliberate attention once reserved for graphical interfaces. The memo still matters—but increasingly, the prompts are the product.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.
