I keep seeing people compare Fable 5 to GPT-5.6 Sol, but that feels like the wrong matchup
If Fable 5 is mostly per-token after July 12, the real comp is Opus 4.8 / GPT-5.6 as workhorse models
Im just tryna figure out who to throw my money to
Anyone tried claude-video by bradautomates?
New-ish GitHub repo with 6.7k stars. Gives Claude a /watch command: download video, extract frames, transcribe, hand it back.
Thinking of testing it on a UFC clip bc timestamps + fast action expose these tools fast
Feels like the agent loop should be:
run β trace β eval β diagnose β patch the system
Not just:
prompt β tool call β cron summary
Lmk if Iβm missing anything obvious here
Used bradautomates' Claude-video on this bc didn't feel like watching an hour long agent vid
It pulled frames + transcript.
Best takeaway wasn't "give your agent a soul."
It was: add stop conditions, traces, and belief updates
Ex-Google engineer just dropped 1-hour course: loops, self-improving AI, memory systems - from scratch:
00:00 - the self-building agent
03:01 - soul.md runs everything
30:16 - RAG memory: pull 20 messages, not 2,000
31:48 - the loop that knows when to stop
35:14 - find the bug, fix the prompt
50:22 - how Claude compresses your memory
1 hour of his guide beats any paid agent course
watch & bookmark - then read Karpathy's loop method below
Practical checklist Iβm pulling from this:
1. What should survive in memory?
2. What does "done" mean?
3. What did it actually inspect?
4. Where did it fail/get stuck?
5. What rule changes next time?
Claude-video was fun here, but still janky on my setup
Pulled the X vid frames cleanly. Transcript needed a workaround, and the full URL flow wanted things I didnβt have configured
Still better than scrubbing an hour long video manually
claude-video successfully ingested the clip and got the KO sequence right.
But timestamp precision was a little off: it put the knockout around 0:20 instead of the correct ~0:14
Useful, but Iβd verify timestamps before trusting them