Claude LFE - Library First Engineering
retained text
I'm a reliability engineer by background (marine + biotech), and I've been building AI products long enough to hit the same wall over and over: Claude nails the prototype, then across a dozen sessions it quietly re-litigates settled decisions, forgets intent, and sprawls into spaghetti. The bottleneck was never capability. It was trust. So over the last while I built a scaffold for Claude Code where the docs are the source of truth, personas are tool-locked (the agent that writes plans can't write code; the one that inspects can't edit production code), and the gates are mechanically enforced with hooks — not "please remember to." It's what I actually use on my own projects; it grew out of a personal project I've been grinding on for a long time. The part I most want torn apart: I built an eval harness that plants defects and measures whether my inspector skills actually *catch* them. I keep seeing guardrail frameworks that assert they work and never show a number. Mine has a scorecard. Two honest questions for this sub: 1. Is the overhead worth it, or am I paying a discipline tax you'd never accept in practice? 2. With Claude Code now shipping native subagents, hooks, and skills — how much of this does Anthropic eventually absorb, and what's left that's mine? Happy to link the repo + the scorecard in a comment. Rip into it.
document fetched 2026-10-13 00:42 · canonical https://reddit.com/r/ClaudeCode/comments/1x3t51p/claude_lfe_library_first_engineering