Flask Review Loop

Flask is the feedback layer for video. Reviewers hit record and talk through the video (voice, camera, screen, drawing on frames) instead of typing; Flask turns the recording into individual timestamped comments with transcripts and summaries, so you can read every piece of feedback as text anchored to a video timestamp. Typed comments arrive the same way.

All access goes through the Flask MCP server (flask), included with this plugin (https://api.flask.do/api/mcp/mcp). If its tools are unavailable, the user needs to authenticate once: tell them to run /mcp, pick flask, and complete the browser sign-in.

The loop

1. Start the upload - uploadfilestart(folderid, filesize, title, content_type).

Find a folder with contents() if you do not have one. The response includes the shareable link.

1. Give the user the link IMMEDIATELY - before uploading any bytes. They can open

it right away and watch the upload and processing happen live. Never wait for "ready" to share the link.

1. Upload the bytes with the exact curl command from the tool's next_step

(a presigned PUT - the Content-Type header is required).

1. Finalize with uploadfilecomplete(asset_id). The video is watchable

instantly via preview playback. Do not poll asset_status; there is nothing to wait for.

1. Listen for feedback - waitforfeedback(asset_id). It blocks up to ~45s and

returns new feedback the moment it is left. Loop it, passing the returned next_since each time, while the user reviews.

1. Read the feedback properly:

- content is the verbatim text of a comment; timestamp is seconds into the video. - Recording items carry recording.transcript (what the reviewer said, sliced to this segment) and recording.aisummary. Treat the transcript as the feedback. - When the transcript references something visually ("this", "here", "that button", "move it there", or describes a drawn arrow/circle), call getannotatedframes(elementid). It returns the actual frames the reviewer was pointing at WITH their drawing rendered in, plus the transcript marked [FRAME N] so you can see exactly what each reference means. This is the reliable way to resolve ambiguous references - prefer it over guessing from words alone. - (Local fallback for full-res, no drawing overlay: extract the frame yourself with ffmpeg -ss <timestamp> -i <file> -frames:v 1 /tmp/frame.png and read the image.)

1. Implement the changes, re-render, and upload the revision as a NEW VERSION of

the same asset: pass versionof: "<assetid>" to uploadfilestart AND uploadfilecomplete. Never create a separate asset for a revision.

1. Share the versionstackurl - it permanently shows the newest version, so the

user keeps one link for the whole iteration.

1. Go back to step 5. Repeat until the user says they are done.

Rules

• The share link exists from step 1. Surfacing it early is the whole point.
• Quote feedback verbatim when reporting it to the user; include each item's url.
• One asset per deliverable, versions for iterations, versionstackurl for sharing.
• Uploads cap at 5GB. Requires edit access on the folder and a paid/trialing Flask

team - if the server returns a subscription or entitlement error, relay it to the user verbatim.