/watch
Analyze a video, screen recording, or screenshot and produce a structured report
$ npx -y skills add mnvsk97/eyeroll --agent claude-codeHow it fires
How this command gets triggered: by you, by Claude, or both.
- Fires itselfClaude auto-loads it when your prompt matches the work.
- You can call itInvoke it directly when you want it.
- Slash command
/watch
Context preview
What this command does when you run it.
Analyze a video, screen recording, or screenshot and produce a structured report
Command definition
watch.mddescription: Analyze a video, screen recording, or screenshot and produce a structured report
argument-hint: '<url-or-path> [--context "..."]'
allowed-tools: Bash, Read, Glob, Grep
Analyze a video or screenshot with eyeroll and present the findings.
Raw arguments: `$ARGUMENTS`
Step 1: Run eyeroll
Codebase context is auto-discovered from CLAUDE.md, AGENTS.md, .cursorrules, etc.
eyeroll watch $ARGUMENTS --verbose
If the user didn't pass `--context` / `-c` but mentioned something about the video in conversation, add it:
eyeroll watch <source> -c "<what the user said>" --verbose
If eyeroll is not installed, tell the user to run `/eyeroll:init`.
Step 2: Present the report
Read the output. First, check the **Metadata** block at the top for quick triage:
- **category** (bug/feature/other), **confidence**, **scope** (in-context/out-of-context), **severity**, **actionable**
If **scope is out-of-context** or **actionable is no**, note this to the user — the video isn't related to the current codebase.
Then look at the **Content Type** and present accordingly:
**For bug reports:** 1. Metadata summary — category, severity, scope, actionable (one line) 2. **What the video shows** — one sentence 3. **The bug** — expected vs actual behavior 4. **Reproduction steps** — from the report 5. **Key evidence** — error messages, URLs, status codes quoted from the report 6. **Suggested fix** — from the Analysis section 7. Offer: "Want me to fix this? I can investigate and raise a PR with `/eyeroll:fix`."
**For tutorials/how-tos:** 1. Metadata summary (one line) 2. **What the video teaches** — one sentence 3. **Key steps** — the workflow demonstrated 4. **Tools/commands used** 5. Offer: "Want me to turn this into a reusable skill? I can create one with the video-to-skill workflow."
**For feature demos:** 1. Metadata summary (one line) 2. **What's being demonstrated** — one sentence 3. **Key capabilities shown** 4. **How it relates to the codebase** (if context available)
**For feature requests:** 1. Metadata summary (one line) 2. **What's being requested** — one sentence 3. **Desired behavior** vs current behavior 4. Offer: "Want me to spec this out or start building it?"
**For any other content:** 1. Metadata summary (one line) 2. **What the video shows** — one sentence 3. **Key takeaways** from the report 4. **Suggested next steps** from the report
Do NOT dump the entire raw report. Summarize it for the user. Suggest the most relevant next action based on content type.
Read more
description: Analyze a video, screen recording, or screenshot and produce a structured report argument-hint: '<url-or-path> [--context "..."]' allowed-tools: Bash, Read, Glob, Grep
Analyze a video or screenshot with eyeroll and present the findings.
Raw arguments: `$ARGUMENTS`
Step 1: Run eyeroll
Codebase context is auto-discovered from CLAUDE.md, AGENTS.md, .cursorrules, etc.
eyeroll watch $ARGUMENTS --verbose
If the user didn't pass `--context` / `-c` but mentioned something about the video in conversation, add it:
eyeroll watch <source> -c "<what the user said>" --verbose
If eyeroll is not installed, tell the user to run `/eyeroll:init`.
Step 2: Present the report
Read the output. First, check the **Metadata** block at the top for quick triage:
- **category** (bug/feature/other), **confidence**, **scope** (in-context/out-of-context), **severity**, **actionable**
If **scope is out-of-context** or **actionable is no**, note this to the user — the video isn't related to the current codebase.
Then look at the **Content Type** and present accordingly:
**For bug reports:** 1. Metadata summary — category, severity, scope, actionable (one line) 2. **What the video shows** — one sentence 3. **The bug** — expected vs actual behavior 4. **Reproduction steps** — from the report 5. **Key evidence** — error messages, URLs, status codes quoted from the report 6. **Suggested fix** — from the Analysis section 7. Offer: "Want me to fix this? I can investigate and raise a PR with `/eyeroll:fix`."
**For tutorials/how-tos:** 1. Metadata summary (one line) 2. **What the video teaches** — one sentence 3. **Key steps** — the workflow demonstrated 4. **Tools/commands used** 5. Offer: "Want me to turn this into a reusable skill? I can create one with the video-to-skill workflow."
**For feature demos:** 1. Metadata summary (one line) 2. **What's being demonstrated** — one sentence 3. **Key capabilities shown** 4. **How it relates to the codebase** (if context available)
**For feature requests:** 1. Metadata summary (one line) 2. **What's being requested** — one sentence 3. **Desired behavior** vs current behavior 4. Offer: "Want me to spec this out or start building it?"
**For any other content:** 1. Metadata summary (one line) 2. **What the video shows** — one sentence 3. **Key takeaways** from the report 4. **Suggested next steps** from the report
Do NOT dump the entire raw report. Summarize it for the user. Suggest the most relevant next action based on content type.
AI eyes that roll through video footage — watch, understand, act. eyeroll is a Claude Code plugin that analyzes screen recordings, Loom videos, YouTube links, and screenshots, then helps coding agents fix bugs, build features, and create skills.
Other commands on eyeroll.
- /api-watch
Analyze a video URL or local file using the eyeroll hosted API.
Open command - /fix
Watch a bug video, diagnose the issue, fix the code, and raise a PR
Open command - /history
List past video analyses from the cache
Open command - /init
Set up eyeroll for this project — pick a backend, install dependencies, and generate codebase context
Open command

