AI Office Hours
A weekly session for Builders to share what's working with AI, get answers to real questions, and see what's new.
What we'll cover:
AI Review
Open Q&A
Try This
AI Review
What did you build?
Open Q&A
How do I use AI to…
Try This
Set up a simple agent that does the work for you.
No prompt engineering, no wasted tokens. A Claude skill is that agent, and Skill Forge is the meta-skill that builds it.
Five steps
- Step 1Identify the job to be done.
- Step 2Write it down in plain language.
- Step 3Install Skill Forge.
- Step 4Forge your process into an agent skill.
- Step 5Put the agent to work.
Identify the job to be done.
Pick one repeatable job you want handled for you. The example here is a video editor: pull this week's Google Meet recordings and cut them into a highlight reel with music, transitions, and a lead-in slide for each clip.
No job in mind yet?
Connect your calendar to Claude, or take a screenshot of your calendar and ask:
“based on my calendar this week, how might we brainstorm some jobs to be done that can be turned into agent skills”
Write it down in plain language.
You don't engineer a prompt. You describe the job once, the way you'd explain it to a new teammate. Claude runs it from there, pulling in only the details it needs, only when it needs them, so context stays lean.
Install Skill Forge.
Add Skill Forge to Claude first. It is the meta-skill that builds the skill for you.
New problem
A workflow that works still lives in one person's head, and it only runs when they run it.
Solution
Forge it into a skill once. Claude runs it on its own after that, and the same .skill file works for anyone on your team.
Forge your process into an agent skill.
Now create the agent skill. Skill Forge budgets Claude's attention in layers, so the skill loads only what a job needs, and walks you through six steps:
- 1
Discover
Targeted questions about how you actually work: the tasks, real requests, what good looks like, and the edge cases.
- 2
Architect
Map each example to where it belongs: procedures into instructions, repeated code into scripts, decision docs into references, templates into assets.
- 3
Initialize
A scaffold script generates the skill's skeleton and folders.
- 4
Build
Write and test the pieces, then the SKILL.md. Most of the care goes into the description, because that is the trigger Claude reads to run the skill on its own.
- 5
Validate and score
Check the structure, then score it so the skill runs reliably and reads only the context it needs.
- 6
Package
A script zips the skill into a .skill file.
“How might we create a new Claude skill that edits video — pull specific recent recordings from Google Meet, then use ffmpeg to build a highlight video with a lo-fi hip hop soundtrack at low volume. Use transitions and animations between clips to make it feel smoother, and add a contextual slide before each clip that spells out the WHAT and WHY of its key takeaway, so the viewer understands what's in it for them.”
Put the agent to work.
The skill is your agent now. Run it two ways:
On demand
Ask Claude to run it, or let the skill's description trigger it the moment the job comes up.
On a schedule
Set it to run on a schedule, so the work is done before you ask. Every Friday, this week's highlight reel is ready.