← ClaudeAtlas

making-a-narrated-ab-screenshotlisted

Make a narrated, self-explaining screenshot: capture real browser elements over CDP (an API explorer response, a chat UI reply, a request body), lay them out A|B side by side with captions (what to look at, plain meaning of status codes, verdict, a plain-words line), and save one PNG for the user to verify yes/no. Triggers on: narrated screenshot, add narration to the screenshot, caption the screenshot, annotate screenshot for demo, A vs B side by side image, burn in captions, demo slide from screenshots, make it self-explanatory.
sblattj/agent-chandlery · ★ 0 · AI & Automation · score 72
Install: claude install-skill sblattj/agent-chandlery
# Narrated screenshot (A|B + captions) ## More triggers narrated-screenshot (old name), take the screenshot for me, explain what I'm looking at in the image, before/after image for the PR, evidence image One PNG per claim that explains itself: header (claim), step 1 (what was sent to both), A column, B column (each with a yellow "Look for" caption), green verdict, and a one-line plain-words translation. ## Do this 1. Capture each side's raw elements over CDP (`take_screenshot {selector, savePath}`), decluttered first with `evaluate_script` (Swagger UI: hide the `Response headers` / `Request duration` h5 plus its next sibling and `.download-contents`; capture `.live-responses-table` and `#operations-<Tag>-<op> textarea.body-param__text`). 2. `N=${CLAUDE_SKILL_DIR:-${SKILL_DIR:-~/.claude/skills/making-a-narrated-ab-screenshot}}/narr.sh; $N new <workdir> <name>`, then fill every `{{...}}` in `<name>.html` (images are relative to workdir). 3. Open it in a tab labeled `narr`, read `document.getElementById('shot').offsetHeight` and `devicePixelRatio`, and capture selector `#shot`. Crop with `$N crop <in.png> <out.png> 1800 <height>` only in the glitch case below. Read the PNG before showing it, then `open` it in Preview. ## Why it's built this way (friction mined) - ALWAYS `select_page` the tab before driving it (and before switching sides) so the user watches it happen in the foreground instead of hunting for the tab. Users ask for this explicitly. - No PIL / ImageMagick