read-image

Generate textual descriptions of one or more images when the current session model (e.g. opencode/big-pickle) has no image input, by delegating to a vision-capable model on OpenRouter via opencode run with --file. Use when the user drops photos into imgs/, asks "can you read/describe these images", or an image needs a description for an article, portfolio, infographic, or video pipeline. Related terms: 看图, 描述图片, image description, vision model, glm-4.6v, --file.

read-image — describe images via a vision model

When to use

Use when a task needs to know what an image contains but the current session model cannot see images. Typical triggers: the user places photos in imgs/ or a working folder and asks for a description, or the pipeline needs per-image captions/descriptions (articles, infographics, video shot lists).

How it works

Run a one-off opencode run against a vision-capable model, passing the image(s) with --file. The child process reads the image and returns text; the main session keeps working with that text.

Command pattern

Message comes FIRST; the file flag comes after. --file <path> (space form) or --file=<path> (equals form) both work. If --file precedes the message, opencode treats the message as a filename and errors with File not found: <message>.

bash

opencode run "<prompt>" --model openrouter/z-ai/glm-4.6v --file imgs/260816-mexico-01.png

For several images, either loop over files with one opencode run each (answers stay short and per-file) or pass multiple --file flags in one call for a combined read.

Choosing a vision model

Known-good on OpenRouter: openrouter/z-ai/glm-4.6v. When picking another, prefer a model whose name suggests vision support and verify it accepts image attachments (a model with no image input returns a text-only response or errors). The --file flag is supported by opencode run regardless of the model; vision is the model's responsibility.

Prompting tips

Ask for a factual, neutral description of what is visibly present (subjects, objects, colors, composition, any visible text), not interpretation. When the output feeds a video pipeline, additionally ask whether the image contains recognizable human faces — Seedance rejects frames containing real people, so this filter avoids wasted submissions. Note explicitly in the result that descriptions are AI-generated and should be spot-checked by the user for factual mistakes. The face answer feeds the video-gen skill's pre-render face-scan gate: any frame flagged with faces is blocked and swapped for an alternate before a single dollar is spent.

Output handling

For a set of images, instruct the child to reply with a labeled list (e.g. 01: ..., 02: ...) so entries can be pasted straight into the working doc. Save the descriptions to the working file (script, article, notes) as the source for downstream prompts.

Narrative ordering (after the descriptions)

When a set of images must tell a story (essay, slideshow, video), do not trust capture order or preference order — arrange the images so each one motivates the next. A dependable generic arc is a narrowing-then-widening spiral that starts wide and objective and ends on life, not on the theme itself:

1. Establish — a wide, readable scene that locates the viewer (a street, a room, a skyline).
2. State the theme — the most explicit symbol or visible sign in the set (a mural, a title text, an obvious motif).
3. Zoom to the object — the same idea in a small, touchable artifact.
4. Go private — an intimate, sheltered space that "houses" the theme.
5. Go spiritual — a church, temple, or ritual space that places the theme in a larger frame.
6. Introduce people — living subjects doing everyday things; put life beside the theme.
7. End on celebration — festive energy, joy, the theme embraced by the living.

Present the result as a numbered recommendation (order → image → role in the story), keep any image a downstream pipeline rejects (e.g. faces) as a labeled alternate rather than deleting it, and re-tune the order to the actual medium's beat requirements (essay sections, shot list, chapter titles). The video-gen skill's pre-render narrative-order gate consumes this recommendation: render in story order on the first run, never in capture or preference order.

Gotchas

• File not found: <message> — the message was placed after --file. Reorder so the message is first.
• The main model cannot verify the image content itself; treat vision output as untrusted until the user confirms.
• If the vision model fails to load the image (binary/extension issues), convert to PNG first (ImageMagick convert) — PNG attachments are the most reliably supported.