SDD vs SDD¶
March 7, 2026 · LinkedIn
11 reactions · 3 comments · 1,417 views
SDD meant one thing last year. It means something different now.
In 2025, Spec-Driven Development was the industry's answer to vibe coding. Write a spec. Hand it to the agent. Get structured output.Responsible, repeatable, better than chaos.
But every session starts from zero. The agent remembers nothing. Tuesday's hard-won lessons are gone by Wednesday.
Skill-Driven Development keeps the same acronym and drops the ceiling. Instead of writing specs that describe what to build, you build what the agent knows. Skills persist. Knowledge compounds. Session 40 is faster than session 4.
We tested this on several projects now. We see 10-40x consistent productivity gains, even on large and complex codebases.. The interesting part was not the speed. It was that the late days were more productive than the early ones. In most projects, the opposite happens.
Same acronym. Different bet. One bets on documentation. The other bets on accumulated knowledge.
Full writeup on the blog: https://lnkd.in/ek4eYPZu
Discussion¶
Same observation (although not as systematic implementation) in my projects. Especially noticeable in large projects.: Same observation (although not as systematic implementation) in my projects. Especially noticeable in large projects.
Totto ↩: Sergiy Yevtushenko Exactly what I keep hearing from people working at scale. The individual session is usually fine — it's session 20 on the same codebase where you feel the compounding effect (or lack of it).
How are you handling cross-session continuity right now in your projects? Curious whether it's a tooling problem or a process problem for you.
Some time ago I realized that the simplest approach is to use CC as an orchestrator for skills/agents and just continue the same session and preserve high-level context. The situation is getting better and better as Claude Code evolves. The addition of background agents and team mode greatly reduced context dilution and made the creation of complex, very thorough skills viable. With the recent extension of the model context window, things work even smoother. Interesting observation: the better the model behaves, the less I need additional tools. Half a year ago I was obsessed with having MCP for everything; now I have none. I've created a small command-line helper inspired by your tools (https://github.com/siy/ndx) and using it in every project, but that's it.: Some time ago I realized that the simplest approach is to use CC as an orchestrator for skills/agents and just continue the same session and preserve high-level context. The situation is getting better and better as Claude Code evolves. The addition of background agents and team mode greatly reduced context dilution and made the creation of complex, very thorough skills viable. With the recent ...