Capability-aware visual quality

Decide whether visual review is possible

Claim visual review only when image review-sheet succeeds and the current Agent can actually understand the downloaded image. A structured unsupported response stops only visual review; command success, filenames, dimensions, generation state, and an asset reference do not prove visual quality.

Record one honest result:

• completed: every required slide and serious presenter candidate was inspected and resolved;
• partial: only part of the required set was inspectable;
• not_performed: the Agent could not consume visual content.

This is completion evidence, not publication permission. Missing vision alone is not a publish blocker.

Review the complete deck

At canonical image readiness, download ordered review sheets with at most six zero-based slide indexes per command and a new bounded destination:

personwise --account <alias> image review-sheet --project-id <project-id> \
  --slides <indexes> --dest <new-path> --json

Zoom or crop locally when detail is needed. Cover every slide; sampling is insufficient. Keep a ledger with slide index/title, intended visual job, text fidelity, semantics/facts, layout, reference expectation, finding, correction, regeneration identity, and recheck.

Inspect:

• alignment with title, key points, page text, narration, and sources;
• visible text spelling, hierarchy, size, overlap, clipping, and table legibility;
• whether the visual explains the intended mechanism, process, comparison, evidence, or idea;
• deck-level continuity and useful variation without repeated template composition;
• invented numbers, charts, quotations, maps, artifacts, customers, logos, UI, workflow states, or source-looking material;
• unsafe advice, inappropriate imagery, watermarks, and intellectual-property problems;
• approved reference influence and any required subject preservation;
• presenter placement and mobile-scale readability.

Treat icons and composition as claims: a checkmark implies success, a lock implies security, a rising chart implies growth, a dashboard implies observed UI/data, and an archival treatment implies a real artifact. Reject unsupported implications.

Correct and regenerate

Fix factual/content causes first:

1. Read a fresh course snapshot.
2. Apply the smallest aligned revision-bound course update batch.
3. Read the new revision.
4. Put the complete failed subset and concrete per-slide instructions into one image regenerate command.
5. Use bounded run wait/run get until each target is canonically complete.
6. Download a new review sheet for the changed subset and inspect again.

Do not split one failed subset into prompt-lottery calls or regenerate based on imagined defects. After an ambiguous response, read authoritative state before another mutation.

Use local replacements narrowly

Only a human or vision-capable Agent that genuinely reviewed the exact bounded PNG/JPEG/WebP may attach it as a reviewed reference/replacement. Correct the lesson first, verify 4:3 composition and every visible claim, fetch the current revision, upload through the CLI's bounded image command, then request a fresh review sheet. Transfer success alone is not a visual pass.

Review presenter candidates

Use presenter list to establish exact language/voice readiness. With vision, use presenter preview --presenter-id <id> --dest <new-path> --json for serious candidates and inspect observable formality, warmth/authority, energy, expression, clothing, framing, and teaching fit.

These are casting observations, not identity or biography. Appearance does not establish nationality, ethnicity, profession, personality, or history. Select a complete compatible presenter/voice pair.

Non-vision path

When vision is unavailable:

• require canonical completed image state for every slide;
• do not claim text, layout, semantics, variety, or reference influence passed;
• do not regenerate or upload a supposedly reviewed image from fabricated findings;
• use only structured presenter/language/voice compatibility and the deliberate default;
• make no appearance or cultural-fit claim;
• record visual_review_status=not_performed with the real reason;
• continue the authorized workflow from structured state.