// Analyze media files (images, video, audio) using AI vision and audio models with a critical-first lens. Use for output-first verification — render your deliverable, then read_media to see what it actually looks like. Distinguishes fundamental issues from surface-level fixes.
Quick one-shot evaluation of work against quality criteria. Generates criteria, runs the round-evaluator, and presents the verdict. Lighter than /refine — does not auto-improve, just evaluates. Use for pre-PR quality checks or getting a critical assessment.
Launch a MassGen multi-agent run. Multiple LLM backends (codex, gemini, claude, grok) collaborate on a task via voting and consensus. Runs in Docker containers by default.
Run a MassGen-style quality refinement loop. Generates eval criteria, produces/improves answers with subagent help, regression-guards before submission, then evaluates with round-evaluator and trace-analyzer in parallel. Invoke with /refine.
Launch parallel MassGen step mode processes — the lead manages all background tasks directly, tracks answers/votes, detects consensus, and synthesizes the result.
| name | read-media |
| description | Analyze media files (images, video, audio) using AI vision and audio models with a critical-first lens. Use for output-first verification — render your deliverable, then read_media to see what it actually looks like. Distinguishes fundamental issues from surface-level fixes. |
Analyze media files using read_media from the media-tools MCP server. This
is the primary tool for output-first verification — experiencing your work
as a user would, not just reading code.
The vision model is instructed to be a critical reviewer by default:
"You are reviewing this work critically. Be honest about what you see — name specific problems, explain their impact, and distinguish between issues that need a fundamental rethink vs issues that are easy fixes. Don't sugarcoat."
When no prompt is provided, the default is:
"Analyze this {media_type}. What works, what's broken, and what would a demanding user complain about? Be specific and critical."
Responses include a foundation_sound assessment that flags when the
approach itself needs rethinking vs. when incremental fixes are sufficient.
When foundation_sound: false, treat this as a signal to consider
TRANSFORMATIVE changes rather than more polish.
You don't need to add "be critical" to your prompts — it's the default. Focus your prompt on what to look at.
# Analyze a screenshot
read_media(prompt="What flaws or layout issues do you see?", file_paths=["screenshot.png"])
# Compare before and after
read_media(prompt="Compare these two versions — what improved and what regressed?",
file_paths=["v1.png", "v2.png"])
# Check a video
read_media(prompt="Is the animation smooth? Any glitches?", file_paths=["recording.mp4"])
# Verify audio output
read_media(prompt="Is the speech clear and natural?", file_paths=["output.mp3"])
| Parameter | Description | Example |
|---|---|---|
prompt | What to analyze — focus on what to look at | "Evaluate the visual hierarchy and spacing" |
file_paths | List of file paths to analyze together | ["before.png", "after.png"] |
continue_from | conversation_id from a previous call for follow-ups | "conv_abc123" |
max_concurrent | Max parallel analyses (default 4) | 4 |
All files are sent to the model in a SINGLE call — the model sees them side-by-side. This is critical for comparison: if you send files in separate calls, the model cannot compare them.
# CORRECT: both files in one call — model sees both
read_media(prompt="Compare before and after — what improved, what regressed?",
file_paths=["v1.png", "v2.png"])
# WRONG: separate calls — model has no memory of the first
read_media(prompt="Analyze this", file_paths=["v1.png"])
read_media(prompt="Now compare to previous", file_paths=["v2.png"]) # can't compare!
For follow-up questions on the same media, use continue_from:
# First call
result = read_media(prompt="Evaluate the design", file_paths=["page.png"])
# result contains conversation_id
# Follow-up (model remembers the image)
read_media(prompt="Now focus on the typography — is the hierarchy clear?",
continue_from="conv_abc123")
Formats: png, jpg, jpeg, gif, webp, bmp
| Backend | Default Model | API Key |
|---|---|---|
| Google Gemini (priority 1) | gemini-3.1-pro-preview | GOOGLE_API_KEY / GEMINI_API_KEY |
| OpenAI (priority 2) | gpt-5.4 | OPENAI_API_KEY |
| Claude (priority 3) | claude-sonnet-4-5-20250929 | ANTHROPIC_API_KEY |
| Grok (priority 4) | grok-4.20-0309-reasoning | XAI_API_KEY |
| OpenRouter (priority 5) | openai/gpt-5.2 | OPENROUTER_API_KEY |
Formats: mp4, mov, avi, mkv, webm, gif, flv, wmv
| Backend | Default Model | Method | API Key |
|---|---|---|---|
| Google Gemini (priority 1) | gemini-3.1-pro-preview | Native video (no frame extraction) | GOOGLE_API_KEY |
| OpenAI (priority 2) | gpt-5.4 | Frame extraction (8 frames default) | OPENAI_API_KEY |
| Claude (priority 3) | claude-sonnet-4-5-20250929 | Frame extraction | ANTHROPIC_API_KEY |
Formats: mp3, wav, m4a, ogg, flac, aac
| Backend | Default Model | Mode | API Key |
|---|---|---|---|
| Google Gemini (priority 1) | gemini-3.1-pro-preview | Native audio understanding | GOOGLE_API_KEY |
| OpenAI (priority 2) | gpt-4o-audio-preview | Rich analysis (tone, emotion, pacing) | OPENAI_API_KEY |
Backend auto-selects based on available API keys (first available in priority order).
Match evidence to how the output is experienced. Classify by what happens when a user opens it, not by file extension:
| What does it do? | Shallow (incomplete) | Full check (required) |
|---|---|---|
| Stays still (image, PDF, document) | File generates | Render and view every page/section with read_media |
| Moves (animation, video) | Single frame | Record video, review full motion sequence |
| Responds to input (website, app) | Screenshot looks good | Use it — click buttons, navigate, test states; screenshot each state |
| Produces output (script, API) | Runs without error | Test with varied inputs, capture output |
| Makes sound (audio, TTS) | File exists | Listen via read_media — don't just check file exists |
Before concluding something is broken:
Focus on what to look at — the critical lens is automatic:
| Domain | Prompt |
|---|---|
| Website/landing page | "Assess visual hierarchy, spacing, typography, and responsive layout. What feels generic?" |
| Generated image | "Does this match the brief? What's off about composition, color, or detail?" |
| Chart/diagram | "Is the data clearly communicated? What's misleading or hard to read?" |
| Video/animation | "Is the motion smooth and intentional? Any artifacts, jumps, or timing issues?" |
| Audio/TTS | "Is the speech clear, natural, and well-paced? Any distortion or pronunciation errors?" |
| Before/after | "What improved? What regressed? Be specific per dimension." |
# Playwright (recommended)
npx -y playwright@latest screenshot http://localhost:8765 screenshot.png --full-page
# Multiple viewports
npx -y playwright@latest screenshot http://localhost:8765 desktop.png --viewport-size=1440,900
npx -y playwright@latest screenshot http://localhost:8765 mobile.png --viewport-size=375,812
Then: read_media(prompt="Evaluate at desktop and mobile sizes", file_paths=["desktop.png", "mobile.png"])
On session start, .massgen-quality/environment.json is written with availability
status. Check capabilities.has_vision to confirm read_media will work before
attempting calls.