A minute ago you made "plan my day" a standing instruction that works when you're at the keyboard, in your context folder. Now we build something different — a morning briefing that runs without you answering questions, packaged so it works from anywhere. You'll also meet the tool that makes building these easy: Claude's skill-creator skill.
Good instinct. Here's what's different today.
| Triggered by | Runs where | Asks you questions? | |
|---|---|---|---|
| Plan my day (last session) |
You, in this folder | Project CLAUDE.md | Yes |
| Morning briefing skill (today) |
You, from any folder | Its own SKILL.md | No (by design) |
| Scheduled routine (Part 2) |
A schedule | Anthropic's cloud | No (nobody's there) |
A Skill is about packaging — portable, named, reusable. Skills can be interactive or not. We're shaping this one to be non-interactive at runtime on purpose, because in the next session we're going to hand it real data via connectors and put it on a schedule.
One sentence. What should this skill do for you? We've prefilled the morning briefing as our running example — edit it to make it yours.
This is the only writing you have to do. The skill-creator will turn it into a real Skill for you in the next step. Describe the result you want, not the data sources — picture yourself reading the briefing and what it should tell you.
Describe what a perfect run looks like. Concrete — format, length, what it tells you. Skip the word "useful."
One per line. Each one should be specific enough that a stranger could read the output and grade it.
The most important section. What should it never do?
You don't need to re-describe your role, timezone, or tools here — that's already in your context/Me/ folder from the last session. The skill will pull from there automatically.
Claude Code ships with a skill for making skills. Its job: interview you for whatever's missing, write the SKILL.md file, help you test it with realistic prompts, and iterate until it's good. Think of it as a colleague who's built a hundred of these.
Reads your spec and asks only what's still unclear — usually trigger phrases and output edge cases.
Frontmatter with name and description (that's what makes it trigger), instructions in the body.
Runs the skill on realistic prompts, shows you the outputs, iterates until it passes.
The skill-creator is going to ask you things. Answer in your actual voice — "uh, morning thing," "show me today," "what's up." Those real phrases are what go into the description, and the description is what makes the skill trigger when you ask for it later.
A Skill lives or dies on whether Claude picks it up when you want it. The only way to know is to try phrasings you'd actually use. This is your earliest brush with evals — measuring whether the agent works before you trust it.
Use your actual voice. Typos, lowercase, shorthand — all fine. These become your test cases.
The first pass will probably miss something — the description doesn't match one of your phrasings, or the output is too long, or it skipped a guardrail. That's the point of testing. The skill-creator will help you fix it in the next step.
You don't need to open the SKILL.md file yourself. Tell the skill-creator what you want different and it'll rewrite.
The description is too narrow. Tell the skill-creator which phrasing failed and ask it to widen the description.
The instructions are too vague. Tell it which criterion wasn't met and paste what the skill actually produced.
Repeat this loop — tweak, rerun, check — until all three test prompts produce a briefing you'd actually start your day with.
Your Skill exists. Time to prove it works outside this conversation.
The fix is in the description — edit context/Admin/skills/briefing/SKILL.md and widen the phrasings, or ask the skill-creator to do it for you. This is the same iteration loop from Step 4, just on a real session instead of a test run.
You open a fresh Claude Code session, ask for your briefing in your own words, and get back something that passes your success criteria. In the next session we give it real tools and put it on a schedule.
Today you built stage one. Here's the full ladder from Skill → Skill with live data → scheduled routine. Part 2 picks up at stage two and carries through stage three.
Your SKILL.md is a reusable instruction set Claude loads when you invoke it by name. It has judgment. It doesn't yet have live data, and it doesn't run on its own.
A Skill is a recipe. Add connectors and it can shop for its own ingredients. Put it on a schedule and it cooks without being told to start.
Instructions + judgment in one markdown file. You invoke it by name in any Claude Code session and paste in whatever data it needs.
Same Skill, now with live access to Gmail and Calendar. You still invoke it on demand — but it builds the briefing from your real inbox and calendar, not anything you paste in.
Runs without anyone invoking it. Your briefing lands in your inbox at 7:30am — whether your laptop is open or closed.
Want consistent, high-quality output on demand? A Skill is enough.
Want it to work off your real inbox and calendar? Add connectors.
Want it to run without you? Put it on a schedule.