The biggest productivity gain from AI coding tools isn't faster code generation. It's spending more time on design before you write a single prompt.
That's the consensus from a survey of 15 experienced engineers compiled by Gregor Ojstersek, CTO and founder of the Engineering Leadership newsletter. The findings cut against the common assumption that AI assistants are primarily about speed. Instead, the engineers who report the best results are the ones who slow down first.
Spec-Driven Development Beats Ad-Hoc Prompting
The most consistent recommendation across contributors: write detailed specifications and planning documents before handing work to an AI agent. Engineers who feed Claude Code, Cursor, or Codex a well-structured spec get dramatically better output than those who iterate through vague prompts.
Several engineers also recommended codifying your team's standards in files like CLAUDE.md and AGENTS.md - configuration files that tell AI tools how your codebase works, what patterns to follow, and what mistakes to avoid. Think of them as onboarding documents, but for your AI pair programmer.
The practical breakdown of where AI delivers the most value: test generation, documentation, code review, and repetitive glue code. Architectural decisions, merge approvals, and deployment still belong to humans.
Separate Your AI Agents by Goal
One practice that stood out: using different AI agents for conflicting goals. An agent optimized for shipping speed will cut corners a security-focused agent would catch. Rather than asking one tool to balance both, split the work. Let one agent draft code fast, then run a separate agent focused on security review.
This maps to a broader theme in the responses - separating generation from verification. Don't accept the first output. Run multiple review passes focusing on different aspects: correctness, security, performance, and maintainability.
The Tools These Engineers Actually Use
Claude Code was the primary tool for most contributors. Cursor got strong mentions for its repository indexing, which helps AI understand large codebases. GitHub Copilot was cited mainly for PR review rather than code generation. CodeRabbit and Bito appeared as dedicated AI code review tools.
Model Context Protocols (MCPs) - a standard that lets AI tools connect to external services like Figma, Slack, or analytics dashboards - came up as a way to give agents richer context about what you're actually building.
One stat from the survey: teams using proper orchestration platforms reported that 60%+ of their pull requests were agent-generated. That's a high number, but it comes with a caveat the engineers were quick to add. Context management is the bottleneck. Feed an AI tool poor context, and its reasoning degrades fast. The engineers getting 60% agent-generated PRs are investing heavily in how they structure information for their tools, not just which tools they pick.