AI can accelerate software delivery, while engineering quality still depends on evidence, review, security, and maintainability.
I use AI to support analysis and implementation while keeping technical decisions and accountability with people.
Start with an evidence boundary
Before asking for an answer, define what the system is allowed to treat as fact. In a portfolio project, that may be a CV, a LinkedIn profile, project documentation, and the current source code. In an enterprise system, it may be approved requirements, schemas, platform documentation, logs, and test results.
This prevents polished language from turning assumptions into claims. Unknowns should stay visible until a person or authoritative source resolves them.
Separate analysis from action
AI is effective at comparing options, identifying inconsistencies, drafting structured content, and proposing implementation steps. Actions that change production, permissions, external communications, or sensitive data need a clearer control point.
For each consequential change, I record the evidence used, the approval boundary, the deployed state, and any remaining uncertainty. That separation keeps polished suggestions from being mistaken for verified or completed work.
Make the specification executable
Written constraints are most useful when they shape implementation and verification. A design specification can define typography and motion. An architecture specification can define static output and data flow. A security specification can define headers and credential boundaries. A testing specification can describe what must be checked before release.
AI then works inside the same contract as a human contributor. It does not need to infer the project standard from whichever file it opened first.
Keep review proportional to risk
Not every change needs the same ceremony. A wording adjustment and a production authentication change have different failure costs. Review should focus on factual claims, security boundaries, public actions, data handling, and changes that are expensive to reverse.
For code, the practical checks remain familiar: inspect the diff, build the system, test the affected behavior, and verify the deployed result. AI can help perform those checks, but it cannot redefine success after the fact.
Preserve accountability
A useful AI-assisted workflow leaves a record of what changed, why it changed, what source supported the decision, and what remains uncertain. That record matters more than the novelty of the tool.
AI is most useful when it reduces mechanical work and leaves more attention for architecture, risk, communication, and decisions that require experience.
This website is one example of that approach: the spec-driven portfolio platform keeps design, architecture, security, SEO, deployment, and content decisions beside the implementation.