Executable Knowledge: Quality Increases Velocity
UI는 한국어입니다. 글 본문은 아직 영어 또는 중국어만 있습니다.
Executable Knowledge: Quality Increases Velocity
Full deep dive: Executable Knowledge: Why Quality Increases Velocity in AI-Native Engineering (enterprise-ai-design)
Every team argues about shipping fast vs. shipping well, as if the tradeoff were a law of nature. AI-native engineering teams are proving it isn't. The mechanism is first-pass yield: automated quality gates make changes succeed the first time more often, so less capacity goes to rework.
Ship fast today → Bug → Hotfix → Rollback → Rebuild (feels fast, isn't)
Automated tests + review + deploy + verification
→ higher first-pass success → faster delivery (is fast)
Quality increases velocity.
Knowledge Should Become Executable
The traditional knowledge pipeline — senior engineers' experience absorbed slowly by juniors — has always been lossy. Now it's a hard blocker: an AI agent can't read a senior engineer's brain. It can only read what you've written down.
Senior engineer
↓
Encode it: CLAUDE.md → Skills → MCP servers → Prompts
↓
AI executes those standards on every task, every time
As Boris Cherny, creator of Claude Code, puts it: knowledge should become executable. A wiki page influences behavior probabilistically; an encoded rule influences it deterministically.
Example: Self-Enforcing API Standards
Instead of hoping a new engineer remembers all six items in your API checklist, put it in CLAUDE.md:
Every REST API must include:
- JWT authentication
- OpenTelemetry tracing
- Structured logging
- Unit tests (>80% coverage)
- Swagger documentation
- Health check endpoint
The checklist applies on every generation — nobody has to remember it. Go further and make the review itself executable: CLAUDE.md tells the agent to call an Architecture Review MCP, and the workflow becomes a self-correcting loop:
Generate code → Call review MCP → Fix findings → Regenerate → Done
Quality assurance stops being a phase after development and becomes a property of how code gets produced — including code written at 11 PM, when human review quality collapses.
One Infrastructure, Not a Feature List
CLAUDE.md, Skills, MCP, hooks, and memory look like assorted features. Together they're knowledge infrastructure — different answers to "where does encoded team knowledge live, and when does it execute?" The point was never AI replaces engineers; it's turning your best engineers' expertise into infrastructure that applies to every line of code, not just the lines they touch.
The Stack, Generalized
This is the quality-side complement to loop engineering:
Knowledge → Executable knowledge → Automation → Consistency → Quality → Velocity
Velocity is at the bottom — an output. Teams that chase it directly by cutting the layers above destroy the thing that produces it.
The full article — with the four-claim verdict table, the enforcement argument, and the Related reading — is here: Executable Knowledge.