rec — record a renderer flow, then verify it on frames

Drive the app's renderer with playwright-cli, record an annotated video, then extract frames and read them. The annotations (action callouts, chapter cards) are the point: they make the clip self-explaining, so a vague "record that part" becomes a video you both read the same way — no re-explaining each time.

Read the vague ask, then just record it

The user rarely spells out a script. Map the pointing to a take:

They say	You record
"あそこの部分 / that part"	the route they last looked at → one focused interaction there
"この一連の動作 / this sequence"	the multi-step flow just discussed → one continuous take, a chapter per step
"動作確認して録画 / verify + record"	run the interaction AND judge it from the frames — don't stop at "file produced"

Don't interview them for a storyboard — pick the obvious flow, shoot it, show frames, correct on the next take. Decide and record.