Intent-Driven Development: Strategic Context Engineering for AI
Hare Krishna, CEO of Polarizer Technologies, explains how spec-driven development transforms AI coding from tactical prompting to durable, strategic context engineering. This analysis covers the shift from ephemeral plans to persistent specifications, the role of verifiable intent in reducing technical debt, and the cultural implications for enterprise software delivery.
The Shift to Durable Context Engineering
The integration of AI into software development is moving beyond simple prompt engineering toward a more rigorous discipline: intent-driven development. Hare Krishna, CEO of Polarizer Technologies, argues that ephemeral chat-based planning is insufficient for enterprise-scale projects. Instead, organizations must adopt spec-driven development to create durable context that persists over months or years. This approach treats the specification not as a throwaway document, but as the primary source of truth for the system, ensuring that AI agents operate within a consistent, verifiable framework.
Strategic vs. Tactical Context
A critical distinction emerges between tactical and strategic context engineering. Tactical approaches, such as standard plan modes in coding assistants, offer short-term utility but lack longevity. Strategic context engineering, enabled by spec-driven tools like OpenSpec, formalizes the development process into discovery, design, and tasking phases. This structure allows teams to articulate business intent clearly before implementation, reducing the "intent to implementation deviation" that plagues probabilistic AI models. By maintaining a top-level specification that reflects the current state of the application, teams can verify that new changes do not inadvertently break existing features, a capability known as spec-anchored development.
Operational and Cultural Implications
The adoption of these practices requires a cultural shift from viewing AI as a tool to treating it as a collaborative partner. Teams must focus on articulating intent and domain knowledge rather than merely issuing commands. This shift impacts role ratios and workflows, with QA evolving to validate the context harness itself rather than just code outputs. Furthermore, the ability to use different models for different phases—such as using one for architecture and another for adversarial review—enhances output quality. Ultimately, this methodology paves the way for "software dark factories," where agents execute well-defined tasks autonomously, allowing human engineers to focus on high-level ideation and complex problem-solving. The result is a more resilient, scalable, and maintainable software development lifecycle that aligns technical execution with business goals.
Key insights
-
Ephemeral chat plans are insufficient for long-term software maintenance; durable specifications are required to retain decision context across sessions and team members. This prevents the AI from re-interpreting requirements from scratch in every new session.
Impact: Reduces technical debt and rework by ensuring consistent interpretation of business requirements over the software lifecycle.
-
Spec-driven development formalizes the workflow into discovery, design, and tasking phases, allowing for independent model selection for each stage. This enables teams to leverage the specific strengths of different AI models for architecture, coding, and review.
Impact: Improves code quality and architectural soundness by matching the right model capabilities to specific development phases.
-
The concept of 'spec-anchored' development ensures that the specification and code remain in lockstep, providing a verifiable source of truth. This allows for automated detection of regressions where new features might inadvertently break existing functionality.
Impact: Enhances system reliability and maintainability by providing a continuous verification mechanism against the intended system state.
-
Cultural adoption of AI requires treating the agent as a collaborative partner rather than a command-line tool. Success depends on the team's ability to articulate domain knowledge and intent clearly, rather than relying solely on model intelligence.
Impact: Accelerates team proficiency with AI tools by focusing on communication and context-building skills rather than just technical prompting.
-
The role of QA is shifting from testing individual code artifacts to validating the integrity of the context engineering system. This involves ensuring that the specifications and harnesses used to generate code are robust and free from ambiguity.
Impact: Prevents systemic errors in AI-generated code by addressing root causes in the specification and context layers.
Action items
-
Implement a spec-driven workflow that separates discovery, design, and tasking phases before allowing code generation. Use structured artifacts to capture business intent and technical decisions.
Impact: Aligns AI output with business goals and reduces the frequency of rework due to misinterpreted requirements.
-
Adopt a tool or process that maintains a top-level source of truth specification, ensuring it is updated with every change to reflect the current state of the application.
Impact: Provides a verifiable baseline for regression testing and ensures long-term consistency of the codebase.
-
Define acceptance criteria using structured formats like Gherkin to enable automated verification of AI-generated code against business requirements.
Impact: Increases confidence in AI outputs by providing objective, testable criteria for completion.
-
Experiment with using different AI models for different phases of the development process, such as using a high-reasoning model for architecture and a different model for code review.
Impact: Leverages the specific strengths of various models to improve overall code quality and architectural design.
-
Train teams to focus on articulating domain knowledge and intent clearly, treating AI interactions as collaborative context-building exercises rather than simple command inputs.
Impact: Enhances the effectiveness of AI tools by improving the quality of the context provided to the agents.
Quotes
“For me, the plan mode is tactical context engineering, right? Spec-driven, strategic context engineering. That's how I think of it.”
“I need teammates involved. I need a collaboration pattern there. So that's where I need durable context for that.”
“The biggest marker I see is, do they recognize how big of a shift it is? And do they recognize the cultural aspect of it, right?”