image-read

name	image-read
description	Read (not generate) images via a subagent so the pixel payload stays out of the parent conversation. Dispatches Haiku by default, escalates to Sonnet/Opus when the caller flags ambiguity or needs load-bearing detail.
allowed-tools	Agent

Image Read

Parent Claude is handed an image path (Telegram attachment, screenshot the user pasted, photo dropped into the session). Reading it directly with the Read tool loads ~1000–2000 tokens of image payload into the parent's context and keeps them there for the rest of the session. Over a day of photos that adds up fast.

Instead: dispatch a subagent at a cheap model (Haiku 4.5), have it return a rich text description, and keep the pixel payload out of the parent. Only the text summary lands in main context.

Sibling scope. This skill is about READ/describe only. For image generation see gen-image and image-explore. For hosting images in PRs see gist-image.

When to use

Use when the parent just needs to know what's in an image but is not about to pixel-edit it:

• Telegram attachments the user sent for context (receipts, whiteboards, screenshots of errors)
• Photos the user drops mid-conversation to describe a situation
• Screenshots attached to PR reviews or bug reports
• Any inbound image where a text description is enough to act on

Do NOT use when:

• The parent needs to visually reason over the image itself (e.g. image explicitly asked about by the user as "does this look right?", where a summary strips the evidence)
• The parent is about to modify or diff the image pixels
• The image is something the parent produced and needs to verify end-to-end

Default pattern (Haiku)

Dispatch a subagent with a purpose-aware prompt. The <purpose> slot is the most important knob — it tells the subagent what the caller cares about so the description stays rich where it matters.

Agent(
  description: "Read image",
  subagent_type: "general-purpose",
  model: "claude-haiku-4-5",
  prompt: """
Read the image at <ABSOLUTE_PATH>. Do NOT write anything to disk, do NOT
call any other tools — just read it and describe it.

<purpose>
<PURPOSE_FROM_CALLER>
</purpose>

Return a rich description, not a caption. Someone reading your output later
should be able to answer targeted questions about this image without
re-reading the file. Cover:

- Objects and people (count, positions, relative sizes)
- All visible text, transcribed verbatim (OCR — quote exactly, flag anything illegible)
- Spatial composition (what's in foreground/background, left/right)
- Color palette and lighting, if they carry meaning
- Emotional tone or mood of the scene, if applicable
- Anything unusual, incongruous, or worth flagging

End with a single line:
  confidence: N/10
where N reflects how sure you are that your description captures what the
caller needs given the <purpose> above. Lower the score if text was hard
to read, if the image was low-resolution, or if the purpose asked about
something you couldn't clearly see.
""",
  run_in_background: true
)

Always run_in_background: true. Never block the main thread on an image read — not even when the parent's next decision depends on the image content. The parent keeps working on anything else it can in the meantime and picks the description back up when the task-completion notification arrives. Blocking the main thread on a subagent is the failure mode this skill exists to prevent.

Escalation

Re-dispatch at Sonnet 4.6 or Opus 4.7 when any of these fire:

• Subagent returned confidence: 6/10 or lower
• The returned summary reads as ambiguous or hand-wavy on a point the parent needs (e.g. caller asked "what does the error say" and the summary just says "there's an error message")
• Caller now needs a specific visual detail the Haiku pass didn't cover

To escalate, re-invoke the same template with model: "claude-sonnet-4-6" (or claude-opus-4-7 for load-bearing detail), and tighten <purpose> to name the specific detail you need. Don't re-run Haiku with the same prompt — Haiku already returned its best pass.

Never escalate past what's needed. Most inbound images are fine at Haiku.

Cost signal

Rough order-of-magnitude for a single image read:

• Haiku 4.5 subagent (default): cheap, ~25× cheaper than Opus for the same image
• Sonnet 4.6 subagent (escalate): mid-tier, use when Haiku's confidence was low
• Opus 4.7 subagent (escalate): expensive, only for load-bearing detail
• Parent Read direct (avoid): cheapest per single operation but pixels stay in the parent's context for the rest of the session — the cost isn't the read, it's the carried payload

The tradeoff: one-shot sessions with a single image are fine to Read directly. Anything resembling a multi-turn session with several images should go through this skill.

Example <purpose> slots

Substituted into the template above — the rest of the prompt stays unchanged.

• Telegram home-office photo: "User sent this as context for a conversation about their home office setup. I need what's on the desk, any visible monitors / devices, and the general layout. Transcribe any on-screen or on-paper text."
• PR screenshot of a failing test: "This is a screenshot of a test runner output. I need the exact failing test name, the assertion error verbatim, and any stack-trace frames that mention project source files."
• Receipt: "User wants expenses extracted. I need vendor, date, itemized line items with prices, subtotal, tax, total, and payment method."

Keep <purpose> specific. Vague purpose ("describe the image") gets a vague summary, which forces an escalation that a sharper purpose would have avoided.