---

name: scenario-asset-analysis description: "Use when finished Scenario assets have to give something back: a caption for a dataset or alt text, a reusable style description, a verdict against a brief, a canny, depth, pose, or segmentation control map for the next model, or the asset itself found again by text, tags, or visual similarity and filed into a collection. Keywords: caption, describe, analyze, QA, control map, find similar, reverse image search, semantic search, tag, collection." license: MIT

Scenario Asset Analysis

Overview

Four tools extract captions, style descriptions, analysis, or control maps from existing assets: three fixed-purpose, one open-ended. With search (default toolset) they answer the questions that come after a batch lands, which is where most of the work actually is: is this on brief, what look is this, what does this show, what can the next model condition on, where did last week's approved version go.

None of them are in the default toolset. Get schemas with scenario_tools_search, then run each through the executor matching the returned permission: read uses scenario_tool_execute_read, write uses scenario_tool_execute_write. Use that live result for both estimates and runs; do not infer the lane from an unchanged input asset. Put the tool name in name and its inputs under parameters. Or reconnect with ?toolsets=full. Connection and scope: see the scenario skill in this repo. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Need	Tool	Shape
A caption for one image	asset_caption	details_level: action or action+style
A reusable look off one image	asset_describe	Returns a full description plus a promptable synthesis
Anything else, in your own words	asset_analyze	instruction plus up to 10 images and 10 text_inputs
A conditioning map for another model	asset_detect	modality, one of ten including canny, depth, pose, segmentation

All four bill credits and all four take dry_run: true inside parameters for an estimate first. For OAuth, pass the user-selected team_id and project_id on estimates and runs. A repeated run can create another billed job; an executor lane does not make it free.

The four facts that change the plan

• asset_analyze batches. One call carries up to 10 images against one instruction, so reviewing 200 assets is 20 calls, not 200. num_outputs (1 to 5) is unrelated: it returns several distinct answers to the same instruction, not one answer per image. Ask for a fixed per-image output shape in the instruction so the answers stay parseable.
• asset_detect strips the background by default. remove_background defaults to true, so a depth or pose map comes back with the frame's context already gone. Set it false whenever the map has to cover the whole image.
• It reads a render, not the file. On a vector or layered source it cannot confirm what the file is or report its exact values: one call called a 16-path SVG raster, shifted a #101614 fill to #131A18, and missed an off-palette path. Parse the file for exact color and structure; the tool judges what the render shows.
• Fixed beats flexible when it fits. asset_caption and asset_describe are purpose-built and faster than instructing an LLM to do the same job. Reach for asset_analyze for classification, extraction, comparison, translation, or a verdict against a brief. One carve-out: for a structured pass/warn/fail verdict against a configured brand brief, teams with the Quality Gate add-on have a dedicated tool that stores its verdict and re-reads it free (see scenario-quality-gate); asset_analyze stays the route when the add-on is missing or the brief exists only as prose.

asset_analyze and asset_detect wait up to 180s and then hand back a job_id; carry on with jobs_wait as anywhere else.

Finding an asset again

Retrieval is search with target="assets", and at least one of query, filters, filter, image, or images must be set: an empty call is a 400, not an everything-list.

• By text. query is keyword matching by default; query_semantic_ratio moves it toward meaning (0.5 to 0.8 suits mood queries like "dark medieval atmosphere", 1 is pure semantic). sort_by (say ["createdAt:desc"]) is ignored while that ratio is above 0, so a newest-first list needs keyword mode.
• By similarity. image takes one asset id or image URL and returns lookalikes; images takes {"like": [...], "unlike": [...]} to steer with positive and negative examples (the two fields are mutually exclusive). image_semantic_ratio decides what similar means: 1, the default, matches subject and mood; 0 matches image features, the setting for hunting near-duplicates and crops. Add a query beside it for "like these, but more stylized".
• By structure. filters narrows on kind, tags, model_id, collection_ids, and created_after/created_before, ANDed with everything above.

Worked example: reviewing a batch against a brief

1. Collect the asset ids from the run (jobs_wait returns them).
2. asset_analyze with dry_run: true on the first chunk to price the pass.
3. asset_analyze with images set to 10 ids and an instruction that states the brief and fixes the output: "For each image in order, reply <index>: pass|fail, <reason in under 12 words>, a reason on every line, passes included. Fail anything not centered, not on a plain field, or carrying text." Repeat per chunk. Answers land as text assets (one, or one per image): asset_download them to read the verdicts.
4. asset_describe on the strongest pass. Its promptable synthesis goes straight into the next batch's prompt, which holds the look without a training run (see scenario-consistency); when the synthesis is only a short title, prompt with the description instead.
5. File the result: collection_create, then one collection_add_assets call with every id (it batches past 49 itself and skips ids already filed). asset_add_tags is additive, so tag the failures rather than rebuilding a tag set. The set comes back later with search filters={"collection_ids": [...]}, and its lookalikes with images={"like": [...]}.

Common mistakes

• One asset_analyze call per asset when 10 fit in a call.
• Treating num_outputs as a batch size over images.
• Leaving remove_background at its default on an asset_detect map that must match the source frame.
• Reusing a remembered executor lane: use the permission from fresh scenario_tools_search results, including for dry_run. A lane rejection means refresh discovery and use its executor, not retry the same call.
• Asking asset_analyze to produce an image. It returns text; control maps come from asset_detect and final renders from model_run.