---
title: "I Handed Claude a Loom Instead of a Spec"
url: https://akdarrah.com/posts/i-handed-claude-a-loom-instead-of-a-spec
author: Adam Darrah
published: 2026-10-19T20:45:53Z
updated: 2026-10-19T20:45:53Z
tags: ["ai","tools"]
---

# I Handed Claude a Loom Instead of a Spec

When I make music videos with Claude, I don't describe the edit I want. I hand it a reference clip and say "like this." It works surprisingly well.

It took me far too long to see that product requirements work the same way. Today I recorded a two-minute screen recording: I clicked through a mockup and talked over it, "this part, not that part." I gave the link to my coding agent and asked one question: *is this enough to make a plan?*

It was. Here's what I learned about why it works and how to do it on purpose.

## A narrated walkthrough beats a written spec

A static mockup shows what the screen looks like. A written spec says what it should do, if you have time to write one. A narrated walkthrough captures something both of those miss: **intent**. While you click, you say what matters ("this button is the whole feature"), what's out of scope ("ignore this panel, that's a different project"), and how it should behave over time ("after you save, this turns into an update button").

It's also fast. Two minutes of talking replaces an hour of writing, and you don't have to know in advance which details are important. You'll say them because you can see them.

## What the agent actually does with a video

Be clear about what "watching" means. Today, the agent doesn't watch a video the way a person does. Mine did two things:

1. **Read the transcript.** Most screen-recording tools generate captions automatically. The agent pulled that captions file to get my narration word for word.
2. **Looked at sampled frames.** It downloaded the video, grabbed a still frame every few seconds (ffmpeg works; so does a headless browser seeking through a `<video>` element), tiled them into contact sheets and read them as images.

Transcript plus frames was enough to pick out every screen, modal, button label and confirmation message in the flow. But it's sampling, not watching. Something that flashes past between frames, or that you do without saying, can get missed.

## Pair the video with the live mockup

The recording tells the agent *what* to build. The mockup gives it the exact details: labels, placeholder text, validation messages, empty states.

If your designer sends a link to a clickable prototype, give the agent that too, and tell it to open the link in a real browser (I use the Claude in Chrome extension) and click through every state. If the prototype is a single HTML page, ask it to read the source. Exact strings are sitting right there, along with the logic for what happens on save, apply and update. The plan ends up quoting the real copy instead of paraphrasing it.

If there's a recorded conversation about the feature, add the transcript. That's where the *why* and the explicit "not now" decisions usually live.

## Make it check the video against reality

The step that surprised me most: before planning, I had the agent compare what the video shows with what the code does today. The video is the destination. The code tells you how far away it is. That check turned up things no mockup would show: data the current models don't store, endpoints that don't exist yet, even a test that asserts the new button *doesn't* exist.

Ask for the result as a short table (claim → confirmed / missing / different, with a file and line) and let the agent make the obvious decisions itself. What comes back to you should be the handful of real product questions, not twenty clarifications.

## Tips if you want to try it

- **Keep it short.** Two minutes was plenty for one feature. Record a second video for a second feature.
- **Say scope out loud.** "This, not that" is the most valuable sentence in the recording.
- **Pause on key screens.** Give the frame sampler time to land on them. Open the modal, then wait a beat.
- **Show the end state.** Click all the way through to what the user sees after it works.
- **Bundle your sources.** Video for intent, mockup for exact copy, call transcript for the why.
- **Ask "is this enough?"** Don't ask for a plan straight away. Let the agent tell you what's missing first.

There's a quieter benefit too. When AI does more of the building, it's easy to end up a step removed from the details. Recording a walkthrough keeps you as the author of the intent: the agent works out *how*, but the *what* comes from you, in your own voice.