Follow host evaluation for every requested review.

Automatic execution policy: Read automation guidance first. Creative rules and review findings guide improvements; scores, review/hash/freeze records, CTA choices and intermediate approval steps do not block authorized work. Preserve actual external-action authorization and stop at the artifact the user requested. This policy supersedes older quality-gate wording below.

Inspect existing content

Inspect the requested scope and report evidence. This skill does not edit assets, generate media, approve a board, publish, or spend money. Only use read-only actions of tools listed here. Existing failed or stale reviews remain unchanged. Return findings in the conversation; write a report only when requested. Asset-only requests need no episode.

• Locate the requested channel and existing episode or asset. Reuse the current conversation's target. If several candidates match, list them and clarify the target; never choose another channel's similarly named character. Read the local profile and source files. Missing files are findings, not permission to generate replacements.
• Read plans and reference context using the scope table below. For an episode call mcp__social-flow__storyboard_read with {path: ".../storyboard", level: "full"}. Context is included by default: reviewContext.story, production, and document carry STORY, PRODUCTION and SB_DOC. Use shotId for selectedShot and its neighbour ids; read those neighbours in playback, which follows b-roll insertion and closing-outro order. unscheduledShots flags inserts with no valid playback position. Shot rows include ordered performance, resolved speech, references and media paths. evidenceStatus: "not_verified" is honest metadata, not a failed inspection. Keep the returned source hash. Display positions help the user find shots; stable shotId values identify findings across insertions. Do not assign ids to legacy boards.
• Open actual available reference and output images with the host's native image reader (Claude Read or Codex image viewing). A filename, URL, HTML thumbnail listing, generation receipt, or hash proves no visual detail. Compare pixels with the intended identity, environment, prop state and shot action. Record precisely which images opened.
• Check the whole story in playback order when requested, including silent action, character dialogue, narration, inserts and closing shots. For a single shot, also read its containing scene and immediate neighbours; do not claim whole-story coverage.
• Cross-check planned framing against every claimed visible actor and prop. A close insert cannot also prove a distant witness or a face-worn object outside the frame. Distinguish custody from visibility. Register objects needed to understand a causal link (notebooks, maps, returned payments), and require an explicit location plan for concealed sightlines, locked entrances and safe access. In still-camera plans, split sequential actions or identify the single depicted moment; camera motion does not animate actors. Check the rendered review page for unresolved template placeholders.
• Run read-only structural checks relevant to the scope. Separate machine findings from visual or narrative judgments. Re-read the board hash before reporting; if it changed during inspection, refresh affected context or mark the review stale.
• Report issues, coverage, limitations and the smallest proposed fixes. Stop here.

Scope and context

Read character identity, location assets, prop continuity and character scenarios only for the scopes needed.

For ordered action/dialogue, run the installed plugin's read-only command:

node "$CLAUDE_PLUGIN_ROOT/skills/storyboard/references/check-story.js" "PATH_TO_STORYBOARD" --text

For prop snapshots, run skills/storyboard/references/prop-sheet-cli.js from the installed plugin with the board path and --shot N. N is the current display position from the same board revision, not a stable id. Never pass --apply during inspection. Use mcp__social-flow__storyboard_check for contract findings; a draft without generated assets is a plan review, not a failed visual production or a visually verified result.

Local media boundaries

Read character images directly from the selected channel's local assets/characters/ID/ folder using the host's native image reader. Inspect front, back and face separately and record the actual paths opened. Metadata and hashes alone do not prove visual consistency. If an image is missing or cannot be opened, report visual inspection unavailable for that view and continue with available text. Do not download, sync or generate replacement assets during inspection.

For video distinguish full playback inspected, sampled frames inspected, and not inspected. Sampled frames cannot establish timing, motion, lip-sync or every transition. Read video review when clips exist. If the host cannot inspect playback, state that limitation. Do not automatically listen to or score speech; speech stays not_reviewed unless the user specifically requests it.

Evidence in the answer

Give a verdict for the requested scope, then findings with severity, stable shot/asset id, actual path or quoted action/line, observation, consequence and proposed fix. State coverage as counts (e.g. all 8 planned shots; 5 of 8 generated images opened; no clips). Use verified / issue / unverified / not applicable separately for plan validity, visual agreement and playback. Missing assets or unsupported media stay unverified. Any score must name its rubric and unverified dimensions; do not infer a passing score from a schema check or copy an older score onto new files. Report source hashes for the board and inspected assets when available; no review record is written by this skill.

Acceptance examples: inspection evaluations.