The Biggest AI Productivity Hack? Doing What We Should Have Done All Along
Source: Dev.to
Everyone’s optimizing for AI right now—writing clearer requirements, documenting features properly, structuring code cleanly, maintaining READMEs, and breaking work into small, well‑defined tasks. These practices have always helped AI produce better output, but we already knew they were worthwhile before AI arrived. The missing piece was a fast feedback loop.
Traditional Engineering Practices
- Understand requirements before you start coding.
- Do upfront design and write things down.
- Break work into the smallest possible tasks.
- Document decisions and keep READMEs up to date.
Human beings consistently produce better results when these steps are done well. The problem wasn’t that the practices didn’t work; it was that the feedback loop was too slow—weeks could pass before the benefits became visible.
How AI Accelerates the Feedback Loop
With AI, the cause‑and‑effect relationship is immediate:
- Vague requirements → mediocre AI output.
- Clear, well‑structured input → genuinely useful output.
The speed of this feedback makes the value of good groundwork undeniable to stakeholders who previously couldn’t see the connection.
AI also reduces the overhead of producing the groundwork itself. Documentation, detailed ticket descriptions, and technical write‑ups require less effort, so the “right thing” becomes easier to do.
A Real‑World Example: Kiro
When AWS launched Kiro, its headline feature was “spec‑driven development.” In practice this means:
- Understand the requirements.
- Produce a design.
- Write a task list.
- Start work.
Most mainstream alternatives (e.g., Cursor) focused on generating code directly. Kiro’s insistence on thinking first mirrors what good engineering teams have been preaching for years, and its success as an AI workflow demonstrates that these practices were always the right ones—AI simply proves it faster.
What the Research Says
METR Studies
Mid‑2025 METR study – Experienced developers took 19 % longer to complete tasks with AI, despite believing they were 20 % faster.
Source: METR, “Measuring the Impact of Early‑2025 AI on Experienced Open‑Source Developer Productivity” (July 2025) – metr.orgEarly‑2026 follow‑up attempt – METR acknowledged the need for updated data but struggled to complete the study.
Source: METR, “We are Changing our Developer Productivity Experiment Design” (February 2026) – metr.org
DX / Laura Tacho Research
Survey of 121 000 developers across 450+ companies (Pragmatic Summit, Feb 2026). Measured productivity gains from AI averaged ~10 %, far below the 40‑80 % claims seen on social media.
Source: Laura Tacho / DX, “Measuring Developer Productivity & AI Impact” – presented at Pragmatic Summit, Feb 2026 (reported by ShiftMag)Deeper analysis of 67 000 developers showed a split outcome:
- Some companies cut customer‑facing incidents by 50 %.
- Others saw incidents double.
The differentiator was organizational structure, not the AI itself. Well‑structured teams used AI as a force multiplier; struggling teams had existing problems amplified.
Anthropic Trial
- Controlled trial with junior developers: AI users completed tasks faster but scored 17 % lower on code comprehension. Those who asked questions and built understanding retained higher comprehension, while pure code generation led to “shipping code they didn’t understand.”
Source: Anthropic, “How AI assistance impacts the formation of coding skills” – anthropic.com
Takeaways
- AI is a catalyst, not a silver bullet. It makes the benefits of good engineering practices visible and easier to achieve.
- Productivity gains are modest on average (~10 %) and heavily dependent on team organization and how AI is used.
- Understanding matters: Using AI to augment thinking and ask questions preserves skill development, whereas using it solely for code generation can erode comprehension.
- Invest in the groundwork—clear requirements, design, documentation, and task granularity—whether or not AI is in the mix.
Human written, AI assisted.