Turn your prompt library into Agent Skills: your first SKILL.md (Day 24 stretch, 30-Day Challenge)

The task

Package two entries from your challenge Prompt Library as Agent Skills — folders with a SKILL.md file that Claude Code discovers and loads on demand. This is the bridge between the challenge's two lanes: your Day 2–14 prompt library (a doc you copy from) becomes capabilities your agent applies by itself. Day 24 stretch from the 30-Day AI-Native Challenge.

Before AI

Your prompt library works, but it's manual: find the prompt, paste it, adjust. A skill is the same knowledge packaged so the agent recognizes when to use it — you stop being the retrieval system.

What you'll need

• Claude Code (paid Claude plan or API key — authoring skills is free plain-markdown; running them is where the tool cost lives, same honesty label as Claude Code 101)
• Two proven entries from your Prompt Library — your meeting-notes prompt and email-reply prompt are perfect
• 20 minutes

A plain-words note: SKILL.md is Anthropic's format for packaging agent capabilities, and it's had wider pickup since — but this workflow only assumes Claude Code.

The workflow

1. Create the folder structure. Skills live in .claude/skills/ (project) or ~/.claude/skills/ (everywhere):

mkdir -p ~/.claude/skills/meeting-notes

2. Write the SKILL.md. Frontmatter tells the agent when to reach for it; the body is your prompt, promoted to instructions:

---
name: meeting-notes
description: Use when the user pastes raw meeting notes or asks to structure,
  summarize, or extract decisions/actions from a meeting. Converts messy notes
  into decisions, open questions, risks, and owned next steps.
---

# Meeting notes → structured record

Take the raw notes and produce exactly these sections:

## Decisions
Only things actually DECIDED. Never upgrade a discussion into a decision.

## Open questions
Unresolved items, with an owner if the notes name one.

## Risks raised
Each with who raised it.

## Next steps
Table: action | owner | due date. Missing owner or date → [UNASSIGNED] /
[NO DATE]. Never invent either.

## What I might have missed
Up to 3 clarifying questions about ambiguous items.

Rules: use only what is in the notes. If the notes are too thin to structure,
say so rather than padding.

The description is the trigger — it's how the agent decides this skill applies. Write it like the "use when" line from your prompt library, because that's literally what it is.

3. Repeat for your email-reply prompt (~/.claude/skills/email-reply/SKILL.md) — voice notes, length constraint, and the [NEED FROM ME: ...] no-inventing rule go in the body.

4. Test that they trigger. Fresh Claude Code session:

Here are my notes from today's standup: [paste something messy]

You should see it pick up the meeting-notes skill and produce your exact format — without you naming the skill. If it doesn't trigger, the description doesn't describe what you actually asked (fix the description, not the request).

5. Iterate the same way as your assistant: every miss is an edit to SKILL.md, not a correction in chat. Fixes to the file are permanent.

Verify it worked

Two checks: (1) paste meeting notes with no instructions and get your format — the trigger works; (2) explicitly ask "use the meeting-notes skill on this" and compare — identical output means the skill, not luck, is doing the work.

Troubleshooting

• Never triggers? Description too abstract. Include the concrete signals: "when the user pastes raw meeting notes" beats "helps with meetings."
• Triggers too often? Description too greedy — scope it ("meeting notes," not "any notes or text").
• Output drifts from the format? Move the format from prose into the literal ## headings shown above; structure in the file becomes structure in the output.

Reality check

Two skills won't change your week. Ten will — that's your whole prompt library, self-applying. The compounding move of the challenge is exactly this: every workflow you prove manually becomes a skill you never re-teach. Skills are instructions, not sandboxes: the agent applying them still has whatever file/tool access you granted it, so your Day 26 approval-gate rules still do the real safety work.

Data & security

Skills files are plain text on your machine — keep client names and internal details out of the skill body itself (the notes you paste are governed by your Day 6 playbook, as always).

Going further

Package the Day 26 agent's charter as a skill and you've closed the loop: the capstone agent runs on a capability library you authored. That's the challenge's skills thesis made literal — courses end; a skill library compounds.

Your takeaway

Two SKILL.md files that turn your best prompts into capabilities your agent applies unprompted — the first entries in a library that outlives any course you'll take this month.