Skip to content

/eyeroll:watch

Analyze a video, screen recording, or screenshot and produce a structured report.

Usage

/eyeroll:watch <url-or-path> [--context "..."]

What it does

  1. Downloads the video (if URL) or reads the local file
  2. Extracts key frames (1 per 2 seconds, deduplicated, contrast-enhanced)
  3. Analyzes each frame with the configured vision model
  4. Transcribes audio (if backend supports it and audio is present)
  5. Checks cache for previous analysis of the same source
  6. Synthesizes a structured report using codebase context

The agent then presents a concise summary rather than dumping the raw report.

Arguments

Argument Description
<url-or-path> Video URL (YouTube, Loom, etc.) or local file path
--context "..." Additional context about the video (what it shows, what you want done)
--no-context Skip auto-discovery of codebase context (CLAUDE.md, AGENTS.md, etc.)
--no-cost Suppress cost estimate output
--min-audio-confidence Minimum confidence for Whisper audio segments (0.0-1.0, default 0.4)
--scene-threshold Pixel-diff threshold for scene-change frame extraction (default 30.0, 0=fixed interval)

If you mention context in conversation without using the --context flag, the agent will pick it up and pass it along.

Examples

/eyeroll:watch https://loom.com/share/abc123
/eyeroll:watch ./bug-recording.mp4
/eyeroll:watch ./screenshot.png --context "this error shows up on the settings page"
/eyeroll:watch https://youtube.com/watch?v=xyz --context "implement this feature"

Report structure

The report first classifies the content type, then adapts its analysis sections accordingly.

Content type detection

eyeroll determines what kind of content the video shows:

Content Type Example Analysis Focus
Bug report Error on screen, broken behavior Expected vs actual, fix directions, search patterns
Feature request Mockup or desired behavior What's requested, implementation notes, acceptance criteria
Question / answer request User asks how something works Direct answer, evidence, follow-up questions
Documentation lookup/update Docs are being read or requested Relevant docs, missing docs, suggested update
Tutorial Step-by-step how-to Process, tools used, skill automation potential
Code review PR diff walkthrough Files reviewed, concerns noted
Feature demo Working feature walkthrough Capabilities shown, codebase relevance
General notes Meeting, brainstorm Key takeaways, action items

Report metadata

Every report starts with a structured metadata block:

intent: bug_report | feature_request | question_answering | documentation_lookup | documentation_update | tutorial_or_howto | code_review | feature_demo | product_feedback | general_notes
category: bug | feature | docs | question | other
confidence: high | medium | low
scope: in-context | out-of-context
repo_guess: <repo name or unknown>
repo_confidence: high | medium | low
severity: critical | moderate | low
actionable: yes | no
handoff_recommended: yes | no
  • intent: specific content classification used to decide whether this is a bug, feature request, question, docs task, tutorial, review, or notes
  • category: broad grouping used by older workflows
  • scope: in-context if the video relates to the current codebase, out-of-context otherwise
  • repo_guess: best repo match from provided codebase context, or unknown when unsure
  • actionable: yes if a coding agent can take concrete action, no if informational only
  • handoff_recommended: yes only when another coding agent should make code, docs, test, or config changes

Common sections

Every report includes:

Section Content
Metadata Intent, category, confidence, scope, repo guess, handoff recommendation
Content Type Classification
Summary One-sentence description of what the video shows
What's Happening Step-by-step walkthrough
Key Details Exact text from screen -- errors, URLs, status codes
Audio/Narration What the person said (or "silent recording")
Observations What works, what's broken, what's demonstrated
Environment Clues Browser, OS, URLs, version numbers
Analysis Adapted per content type (see above)
Suggested Next Steps Context-aware recommendations
Agent Handoff Included only when a code/docs/test/config change is recommended
Clarifying Questions Only if something is genuinely unclear

Evidence confidence tiers (bug reports)

For bug reports, the Analysis section categorizes every claim:

  • Visible in recording -- directly observed (error messages, URLs, UI state)
  • Informed by codebase context -- references files from .eyeroll/context.md
  • Hypothesis (not confirmed) -- educated guesses about root cause

This prevents the agent from hallucinating file paths or treating guesses as facts.

Caching behavior

The same video analyzed with the same backend produces a cache hit. Only the synthesis step re-runs, using your new --context. This means:

/eyeroll:watch video.mp4                          # full analysis (~15s)
/eyeroll:watch video.mp4 --context "new angle"    # instant, cached frames
/eyeroll:watch video.mp4 --no-cache               # force fresh analysis

CLI equivalent

eyeroll watch <source> [options]

See CLI Reference for all flags.