Where to Invest Your Time with AI Coding Tools
Should you spend time crafting AI tool configurations, or building things with them? A framework for deciding.
The Configuration Trap
Engineers love optimizing their tools. With AI coding tools like Claude Code, there’s a growing culture around crafting elaborate skills, hooks, and prompt configurations. It feels productive — but is it?
The question isn’t whether these configurations are useful. They can be. The question is whether the time spent building and maintaining them pays off compared to the alternative: just using the tool to build things.
Skills vs. CLAUDE.md — A Key Distinction
There’s a meaningful difference between two types of configuration:
Project facts and rules (CLAUDE.md):
- Directory structure
- Commit message conventions
- Naming patterns
- Team-specific workflows
These describe what is true about the project. They rarely need updating because of the AI model — only when the project itself changes.
Behavioral instructions (skills, hooks):
- Step-by-step reasoning guidance
- Output format enforcement
- Workaround patterns for model weaknesses
These describe how to steer the model. They break when the model changes, because they’re coupled to the model’s current capabilities and quirks.
Why Skills Have a Short Shelf Life
Two forces erode the value of hand-crafted skills over time:
1. Model evolution requires re-tuning
A prompt engineered for today’s model may be unnecessary or counterproductive for the next one. Workarounds for weaknesses become dead weight when those weaknesses are fixed. The maintenance cost is ongoing and unpredictable.
2. Platform absorption
Patterns that many users implement as custom skills get noticed by platform teams and built into the product. This is the same cycle we’ve seen with jQuery → native DOM APIs, webpack plugins → browser standards, and countless other tools. The custom solution becomes redundant.
A Better Framework
Instead of asking “how should I configure my AI tools?”, ask:
- Is this a repeated friction point? If you hit the same issue weekly, a lightweight solution (a prompt template, a CLAUDE.md rule) is worth it.
- Is this a project fact or a model workaround? Facts are stable. Workarounds are temporary. Invest accordingly.
- Would this time be better spent building? Every hour spent on configuration is an hour not spent on output. The experience of shipping things with AI — learning when to delegate, when to intervene, how to iterate — is the durable skill.
What Actually Compounds
The engineers who will benefit most from AI tools in the long run aren’t the ones with the most sophisticated configurations. They’re the ones who have shipped the most things with AI — because that builds judgment that transfers across tools, models, and platforms.
Configuration is maintenance. Output is compound interest.