Turtles

Brown belt at many things, black belt at none.

[Turtles @turtlesfr]

@elonmusk Badass

[Turtles @turtlesfr]

and @AnthropicAI never answered me.

[Turtles @turtlesfr]

Claude Design is a cool tool but it lost an entire file with plenty of mockups, twice. Don't save your work there. Export it somewhere else. Treat #ClaudeDesign as an in-between tool until they make it stable.

[Turtles @turtlesfr]

@oo_kk_aa So good

[okazakitomohiro @oo_kk_aa]

ニャッキの伊藤有壱さんにお声掛け頂き、コマ撮りの展覧会に一作家として参加しています。私はコマ撮り分野ではない場所から活動をはじめて、デザインの視点でのコマ撮りに取り組んできましたが、今回初めてコマ撮り界の本丸の方々とご一緒でき嬉しいです。今6年目のマッチ撮影素材等を展示しています

[Turtles @turtlesfr]

Use Assembly, Whisper, Deepgram... depending on where you have a funcitonal API key already.

[Turtles @turtlesfr]

These webinars you half listen to while you are doing something else? Turn them into easily digestible html webpage so you can extract whatever you need in 5 minutes. Record the webinar (OBS if needed) then use this prompt

-----------------
# Webinar Recording → Slide-by-Slide Static Site

I have a webinar video recording at `<PATH_TO_VIDEO>`. I want you to turn it into a navigable static webpage plus a condensed Markdown handover.

Do all of the following:

## 1. Transcribe the audio properly

- Extract a compressed audio track from the video with `ffmpeg` (mono 96 kbps mp3 is plenty for speech).
- Send it to AssemblyAI for transcription. Use the best model available, enable speaker diarization, and pass a custom-vocabulary list of domain terms, product names, and proper nouns that are likely to be mis-heard — derive that list from any title/description/context I provide (or ask me for one if needed).
- If any existing transcript was provided, treat it as a hint only; the AssemblyAI output is the source of truth.

## 2. Extract unique presentation slides from the video

- Run `ffmpeg`'s scene-change filter on the video (`select='gt(scene,THRESHOLD)'`, start around 0.15) with `metadata=print` so each extracted frame has a timestamp.
- Read the extracted images and visually classify them: a frame is a real slide only if the slide content is materially different from the previous kept slide (target ~90% visual difference at the slide-content level). Discard frames that are duplicates of the same slide with only presenter-video / overlay / cursor changes, and discard non-slide frames (zoom UI panels, accidental window switches, presenter close-ups, empty screens).
- When scene detection misses long stretches (e.g. text-only slides with no transition trigger), probe additional frames at intermediate timestamps and visually check for slides you skipped. Iterate until the deck timeline is complete.
- Crop each canonical slide to remove conferencing UI (top bar, side panels, bottom controls) so the slide content fills the frame cleanly. Save the final set as `slides/NN_short_name.jpg`.

## 3. Reconcile slides with the transcript

- For each canonical slide, determine its `[start_time, end_time]` window in the video.
- From the transcript, pull the speech that falls inside that window. Use that as the per-slide transcript excerpt.
- Distill the speaker's content for that slide into a short, scannable list of key facts (the actual takeaways, not a paraphrase of every sentence). Flag any discrepancies between what's on the slide and what the speaker says — don't silently reconcile them.

## 4. Build a static HTML page (`index.html`)

- Single self-contained file, embedded CSS, dark theme, responsive.
- Sticky left sidebar with a clickable table of contents listing every slide.
- Main column: one section per slide. Each section shows the cropped slide image, its timestamp range, a "Key facts" panel (condensed bullets), and a "Transcript" panel (the speech excerpt with speaker labels).
- Detect the Q&A portion at the end and split each question into its own card: question text, asker name if mentioned, condensed answer, and a topic tag (e.g. pricing / docs / technical / eligibility / process — adapt tags to the actual content). Put the Q&A section below the slides with a prominent "Jump to Q&A" link in the sidebar.

## 5. Write a condensed handover Markdown (`webinar_summary.md`)

- Designed to be pasted into another LLM's context as standalone background.
- Cover: what the product/topic is, status (beta/GA/etc.), prerequisites, the step-by-step flow, the technical contract (APIs, webhooks, data models), limitations, best practices, docs/support pointers, the full Q&A, a glossary of acronyms, and any single critical contact / link / command worth surfacing at the end.
- Flag the same discrepancies you flagged in the HTML.

## 6. Keep deliverables clean

Final tree should be:

```
index.html
webinar_summary.md
full_transcript.txt # AssemblyAI output with timestamps
slides/ # cropped slide images used by index.html
<original video> # left in place
```

Delete intermediate audio files, probe frames, scratch logs, and vocab files when done.

---

**Working style:** work in parallel where you can — kick off the AssemblyAI transcription in the background while you start scene-extracting frames. Use a task list to track progress across the steps. Show me each unique slide as you classify it so I can sanity-check before you commit to the final deck.

[Turtles @turtlesfr]

@tamashenning @steipete Pure curiosity: what is the "SSO tax" about?

[Turtles @turtlesfr]

Asking questions on X when you have very little followers feels lonely and useless. Reddit is much more useful for that. I wish X had a way to actually push genuine questions in front of non followers somehow.

[Turtles @turtlesfr]

I worked a couple of years on a digger. On my second job, I ended up on a 50T machine doing some very repetitive task, on top of a 10m high pile of sand. At the end of the day, the boss told me to take the machine down to fill it up. 10m, 45º slope.... you have to lean on the bucket and literrally do some controled slide down the hill..... terrifying. But the scarier was to get back up! You have to grab as far as possible with the bucket to pull yourself up, moving up, pulling the arm and keeping the machine parallel to the pile of sand... everything synchronized. And halfway up the hill I had to release the arm and extend it again further as fast as possible because the digger would start slidding down again... doing all this while you are at 45º, feeling the whole thing will tip.

Scary when you do it for the first time... But these machines are so fun to work with!

[Turtles @turtlesfr]

I want to try #composer 2.5 through cursor cli. Can I pay per use without a paid plan?

[Turtles @turtlesfr]

@signulll @elonmusk @AnthropicAI @SpaceX If they beg for it

[Turtles @turtlesfr]

@jon_delorraine Monde de timbrés. Cet ado a fait un truc génial et il ne peut qu'être fier de ce qu'il a fait. Je ne vois pas comment on peut se préoccuper de la protection de l'identité d'un pédophile à ce stade...

[Turtles @turtlesfr]

@brivael Congrats on getting into YC. Working my ass off to get there too someday.

[Turtles @turtlesfr]

@levelsio Definitely getting slower... I always had about 5 sessions running in parallel. A couple of months ago I was being limited by my own time to read and answer. Now I find myself waiting much more... and ending up here on X! If this continues I will have to change, sadly.

[Turtles @turtlesfr]

@SpectateursFr Bravo à la personne qui l'a piégé ! Quel beau passe temps de réussir à exposer des sacs à merde comme lui

[Turtles @turtlesfr]

@claudeai Please add support to paste images in there

[Turtles @turtlesfr]

@bcherny I frequently copy paste screenshots. It doesn't seem to be working with "claude agents". Is there a trick to quickly paste a screenshot in there?

[Turtles @turtlesfr]

Any good resources of examples of prompt injections? I want to test my ai chatbot the best I can.

[Turtles @turtlesfr]

@elonmusk @Tesla Black magic