Skip to main content
You’ll learn: How to treat CLAUDE.md and hooks as a system that improves with every session — and why feedback loops are the primary skill of agentic engineering.

​
The Discipline

When Claude makes a mistake, you have two choices:
  1. Fix the instance — correct the output, move on
  2. Fix the system — trace why Claude made the mistake, then patch CLAUDE.md or hooks so it can’t happen again
Option 2 is slower in the moment but compounds over time. After a few weeks of consistent system-level fixes, Claude stops making the same classes of mistakes.

​
Trace, Patch, Never Repeat

The workflow:
  1. Notice — Claude uses relative imports when your project uses @/ path aliases
  2. Trace — CLAUDE.md says nothing about import conventions
  3. Patch — Add to CLAUDE.md: "Always use @/ path aliases for imports. Never use relative paths (../) for cross-directory imports."
  4. Verify — Future sessions follow the convention
Every correction is a signal. The question isn’t “how do I fix this output?” but “what’s missing from my system that let this happen?”

​
What to Patch Where

SignalPatch targetExample
Claude keeps making the same formatting mistakeCLAUDE.md convention”Use single quotes for strings”
Claude does something dangerousHook (hard block)Block git push --force via PreToolUse hook
Claude’s approach to a task type is consistently wrongSkillCreate a skill for database migrations with your patterns
Claude forgets cross-cutting concerns.claude/rules/ file scoped to the relevant pathTesting requirements for src/api/
CLAUDE.md handles conventions. Hooks enforce safety. Skills encode complex workflows. Rules scope context to specific areas. Use the right tool for the signal.

​
Feedback Loops as Primary Skill

The core skill of agentic engineering isn’t prompting — it’s designing tight validation loops. The tighter the loop, the faster Claude self-corrects.
Loop tightnessExampleFeedback speed
TightestPostToolUse hook runs linter after every editSeconds
Tight”Run npm test after each change” in the promptMinutes
LooseReview at the end of a sessionCould be too late
The tightest loops are automated (hooks). The next tightest are embedded in the task prompt. The loosest require your attention — reserve those for judgment calls that can’t be automated.

​
The Compounding Effect

A well-maintained system produces a flywheel: 
Better CLAUDE.md → fewer mistakes → less correction time →
more time for system improvements → even better CLAUDE.md
Engineers who invest in system evolution consistently report that Claude “gets better” over weeks. The model isn’t improving — their system is. This is the self-improving loop at the practice level.
← Prev: Robots-First · Next: Distribution →