While Elon Musk talks about how by the end of the year “code will be unnecessary,” it becomes clear: moving from simple coding to clear task and procedure formulation is essential. This has always been challenging.
Both OpenAI and Anthropic are publishing guides on how to think about skills correctly — how to describe, structure, and create them.
Key recommendations:
Skill structure:
- The description should include WHAT the skill does + WHEN to use it
- Add negative examples (“DO NOT use when...”) — this improves accuracy by ~20%
- Embed templates and examples directly into the skill, not in the system prompt
How to write instructions:
- Be specific instead of abstract: not “validate data,” but “run python scripts/validate.py --input {filename}”
- Document error handling explicitly
- Keep the main file compact — move details to separate files
Testing:
- The skill should trigger on 90% of relevant queries
- It should not trigger on irrelevant ones
- Iterate on a complex task until it works, then extract the approach into the skill
The main takeaway: it’s time to learn how to automate non-trivial processes. Trivial ones will be automated without you by tomorrow.
I recommend checking out the detailed instructions from both companies:
- OpenAI: developers.openai.com/blog/skills-shell-tips/
- Anthropic: claude.com/blog/complete-guide-to-building-skills-for-claude
