How to Use Claude Skills 2.0 Better than 99% of People

TL;DR

Skills 2.0 adds built-in automated evals so you can test and optimize Claude skills far faster than manual iteration.

Key Points

• 1.What Skills 2.0 adds: The skill creator skill now includes an Eval Viewer agent and benchmarking scripts that run multiple test variations simultaneously and score them automatically against your defined criteria.
• 2.How to prompt evals correctly: Always specify one optimization target (e.g., copywriting style), define explicit scoring criteria (e.g., m-dashes, personal stories, word count, reference file match), and set the number of test variations — the creator showed 5 variations on one YouTube video.
• 3.Why specificity matters: A vague prompt like "run some tests" lets Claude invent its own criteria, producing useless results. Precise criteria (style match, word count, personal stories) revealed 2/5 failures immediately and gave actionable optimization targets.
• 4.How AB tests work: Claude spins up a second version of the skill (Version B) optimized for your goal (e.g., speed), then runs both on the same input. The original skill used 93,000 tokens in 204 seconds; the optimized version used 77,000 tokens in 160 seconds.
• 5.When to use AB tests vs. evals: Use evals to fix a broken or underperforming skill; use AB tests only once a skill already works well, to push it from good to great — e.g., testing whether removing a reference file improves or harms output quality.
• 6.Skill structure that improves outputs: Include a trigger definition, main goal, required connectors (e.g., YouTube transcript MCP), reference files (voice, ICP, writing framework, newsletter examples), a step-by-step process with human-in-the-loop checkpoints, and a self-learning "progressive updates" rule so the skill auto-updates from feedback.