#7dayAISchallenge — Day 3: Skills 👩🏻‍🍳

Spent today learning about Claude Skills. Built one following the day 3 instructions, then rebuilt it through Anthropic's skill-creator to see what it does under the hood. Sharing the stuff that wasn't obvious to me going in. 👇

1. The description is the trigger contract. Not the body. 🎯

This was the biggest unlock for me. Claude only sees a skill's name + description at conversation start. The body isn't loaded until after the skill is picked.

Which means: a strong body with a vague description will rarely fire. A simpler body with a sharp description gets used reliably.

2. Some description-writing rules that actually help ✍️

Start with e.g. "Use when..." frames it as user intent, not capability
List the actual phrases users type (synonyms matter)
Add negative space: "even if they don't explicitly say X". This catches paraphrased asks
Keep it short. Descriptions are loaded every turn for every skill, so tokens add up
Skip the marketing. "Best-in-class image generator" doesn't help Claude match anything

3. Two skills with overlapping descriptions = unstable triggering ⚠️

If two skills can both match the same prompt, Claude just picks one and it's not consistent across conversations. One skill should own a trigger surface. If you've got a v1 and v2 hanging around, delete v1.

4. The body plays by different rules 📋

Once a skill fires, the body is the playbook. Different rules from the description:

Write imperatively not descriptive: "Confirm the prompt" not "This skill confirms prompts". Claude is reading the body to decide what to do. "This skill confirms prompts" is a description of behavior where Claude has to translate it into an action. "Confirm the prompt" already is the action. Less ambiguity, less room to interpret it as flavor text.
Rule of thumb: imperative when the model should follow the text (skill bodies, system prompts, user prompts). Descriptive when the model is deciding whether to use the text (skill descriptions, tool definitions).
Number the steps. Claude follows them in order
If your reference docs are long, put them in a separate file so the body stays scannable
Spell out failure modes. "If the API returns 429, wait and retry once" beats "handle errors"

5. Skill-creator runs a 4-phase loop 🔄

When you invoke Anthropic's skill-creator (meta skill), it doesn't just spit out a SKILL.md file. It walks you through four phases:

1. Elicit — it asks you questions. What's the skill for? What should trigger it? What are the edge cases? Think of it like a PM doing requirements gathering before any code gets written.

2. Scaffold — it generates the actual files: SKILL.md (with the description and body), plus any supporting files like API specs or reference docs. This goes into a workspace folder, not into your live skills directory yet.

3. Eval — it generates test cases (typically 3 diverse prompts a real user might type) and runs them two ways: once with your new skill loaded, and once without. The "without" run is the baseline — what Claude does naturally. This is how you know if your skill is actually adding value.

4. Optimize — a judge LLM scores both runs against the success criteria you specified. If the skill underperforms, skill-creator rewrites the description or body and runs the evals again.

So it's a workflow, not a generator. A generator would hand you a file once and call it done. A workflow keeps looping until the output measurably works.

0 comments