DEV Community

Keo Fung | FormLM
Keo Fung | FormLM

Posted on

I Let an AI Agent Prep My Workshop. Prep Took 20 Minutes.

#ai #workshop #productivity #mcp

Series: How to Run a Workshop People Remember (7/10)

Workshop prep is the part nobody teaches and everybody dreads: read the run sheet, list the props, check print specs, build the sign-up and feedback forms, write the invite. It's all mechanical work that takes an evening — which is exactly the kind of work an AI agent should be doing while you do the human parts.

The blocker was always that AI couldn't see your materials. It could draft you an email, but it couldn't open your activity folder, know the group size, count the props, or tell you that the A3 poster needs a color printer. So I made the materials machine-readable, and this post shows the workflow — because it generalizes to any prep materials you keep in files.

Make the prep a skill file

Every folder on WorkshopSkill ships with a SKILL.md file — frontmatter (name, description, group size, duration, props) plus runnable prep commands. The point of the format: an AI agent that can read files can discover the activity, install it, and run the prep without you explaining any of it in the chat.

WorkshopSkill's skills page: every folder installs as a skill for AI agents — a discover, install, prep three-step workflow

The three-step workflow, in practice:

  1. Discover. The site publishes a machine-readable index — skills.json, one entry per folder with the bundle URL and a checksum. Point an agent at it and ask "I need a 20-minute pairs activity on communication for 16 people" and it can shortlist from actual metadata, not vibes.
  2. Install. Download the folder (each has a files/SKILL.md at a predictable URL — here's Blind Portrait's, for example). The agent reads the run sheet, the props list and the prep commands the same way you would.
  3. Run the prep. This is where it stops being a gimmick. The agent produces the actual checklist: "print 8 prompt cards poker-spec, A4; print the A3 poster; cut cards; you'll need markers; run the pre measure before doors open." It knows all of this because it's in the folder, not because the model guessed.

The skills-for-agents page documents the workflow for humans, and the skill tutorials walk through it per client and per scenario (28 tutorials at time of writing — Claude, Cursor, and the rest, each with its quirks). There's also a plain guide on AI agent workshop prep for the concept without the tooling.

The forms, via MCP

The prep checklist has always ended with the annoying tail: "build the registration form, build the pre/post measure, build the feedback form." Those are all form-shaped, and form-building is now an MCP-shaped problem.

FormLM ships an open-source CLI and MCP server (the npm package @formlm/cli). If you use Claude, Cursor or any MCP-compatible client, the whole install is:

npx -y @formlm/cli mcp
Enter fullscreen mode Exit fullscreen mode

One command, stdio transport, no API keys to wire up — a client with no token signs in through the chat. Then your agent can build the workshop's forms and assessments as part of the same prep session: describe the pre/post measure once, get the live scored form, the registration page and the feedback form — each with a URL you can drop onto a printed card (part 3) or into the invite email.

So the whole prep conversation with the agent ends up being one message:

"Install the blind-portrait folder from skills.json, give me the prep checklist, then build the registration form and the pre/post measure on FormLM for my 16-person session Thursday."

Twenty minutes later you have: props checklist, print specs, a run sheet summary, and three live form URLs. The evening you would have spent on this is now the evening you spend rehearsing the debrief questions — the part that actually decides whether the workshop is any good.

Honest limits, learned the hard way

  • The agent doesn't run the room. It can tell you to print twelve cards; it can't make the printer cooperate. Prep, yes; facilitation, no.
  • Machine-readable beats clever prompting. The workflow works because SKILL.md and skills.json state facts the model can't hallucinate. When I prep with materials that aren't structured this way, the agent confidently invents group sizes and prop counts — garbage in, confident garbage out.
  • ChatGPT-style drafting still has a seat. For the invite email and the one-page agenda, chatgpt.com or claude.ai with your prep checklist pasted in is plenty. The agent-with-tools route earns its setup cost when you do this every month.
  • Trust, but verify. I read the agent's checklist against the human-readable folder page every time. Not because it's often wrong — because the one time it's wrong is the night before.

Next in the series: the debrief — the beat where the learning actually happens, and the one most workshops skip.

Tools in this post:

*Disclosure: I'm one of the makers — the tools from my own shop named across this series. *

Top comments (0)