Skip to content
AnyCapAnyCap Docs

Review Media with Human Feedback

Use non-blocking annotation when an agent must pause for a human review without holding an interactive command open.

  1. Start a review session.

    Terminal window
    anycap annotate ./draft.png --no-wait

    Save the returned session, url, poll_command, and stop_command. Show the URL to the reviewer.

  2. Wait for human confirmation.

    The reviewer draws rectangles, arrows, points, or freehand notes, optionally records narration, then clicks Done.

  3. Retrieve structured feedback.

    Terminal window
    anycap annotate poll --session <session-id>

  4. Analyze narration when present.

    Terminal window
    anycap actions video-read \
      --file .anycap/annotate/<session-id>/recording.webm \
      --instruction "List requested changes, their locations, and any constraints"

  5. Revise the annotated image.

    Terminal window
    anycap image generate \
      --prompt "#1 remove the object; #2 move the headline right; preserve everything else" \
      --model <model-id> \
      --mode image-to-image \
      --param images=./draft_annotated.png \
      -o draft-v2.png

  6. Stop the local session.

    Terminal window
    anycap annotate stop --session <session-id>