---
name: record-a-task-get-a-skill
description: Turn a narrated screen recording of someone doing a repetitive task into a working Claude skill. Transcribe the walkthrough, find the steps, inputs, and preconditions, ask about the gaps the narrator skipped, and write the SKILL.md. Run it supervised first, then remove the checkpoints. Use when non-technical teammates keep doing the same multi-step chore by hand.
---

# Record a Task, Get a Skill

> **Where this came from.** Distilled by Calvin (Colton Mulligan's Claude) from Marty Reed's demo at AI Open House, Session 001, 45:01 in the recording: https://aiopenhouse.co/sessions/001/?t=2701
> Marty built a working recorder that does this end to end; this file reconstructs the method he explained, not his tool. Marty's walkthrough of the real thing is on the way. Credit Marty if you use it. Shared under CC BY 4.0.

## Who it's for

People who will never write a skill but can show you their job. As Jared Egli put it on the call: it's video vibe coding.

## Step 1: record the task, narrating out loud

The person does the chore once while talking through it: where they click, what they check, why. Marty's example: setting up each client's QuickBooks workspace integration, a tedious, error-prone process he used to do by hand.

## Step 2: transcribe, then give this skill the transcript

## Step 3: build the skill

From the transcript, write down:

- **Name** and a one-line description of when to use it.
- **Inputs** it must ask for (client name, account, dates).
- **Preconditions**: what must already be true (logged in, file exists, access granted).
- **Steps**, numbered, one action each.
- **Checks**: where the narrator verified something, and what "right" looks like.
- **Gaps**: anything the narrator skipped or assumed. List these as questions and ask them before writing the final skill.
- **Improvements**: steps that could be removed, combined, or done by the agent instead.

Write it as a `SKILL.md`: frontmatter (name, description), then the sections above.

## Step 4: run it supervised

The first few runs pause for the human at every check, in a browser they can watch. Seeing it do the work is what wins people over.

## Step 5: take the training wheels off

Once it runs clean, tell the agent: "Remove the checkpoints from this skill and save it." Now it runs by itself.

## Step 6: stack skills into a plugin

Several related skills (say, a QuickBooks entry, then an invoice, then a reminder) can ship together as one plugin.

## Marty's rule underneath it

Do the work through the agent, not around it. When the task runs through Claude, Claude builds the context, and the next time it's just "create these invoices. Want me to send them?"

## A second lesson from the same demo

When you research at scale, break the job into small chunks on a schedule: not "research every state's policy," but state by state, then topic by topic, each on its own recurring job. Big asks produce thin answers; small ones stack into a rich data set.