Founder / Software Architect @ KAM Studio. Building AI systems, software and digital products. Sharing what I build, test and learn.
Joined September 2026
- Tweets57
- Following99
- Followers29
- Likes72
You write what the film should be, briefcut makes it. No editor, no timeline. Any video model or your footage, voiceover, music, every sound on its exact frame, with a QA report. The brief asked for three hammer taps, the model made seven, briefcut found all seven. Open source.
Repo, full build log of this film and a demo that runs without any API key:
github.com/cagrikacmaz/brief…
npx github:cagrikacmaz/briefcut demo
I gave Gemini 4 Argon my live product and let it ship to production twice. I wrote the brief and the constraints, it did the audit, the fixes, the tests and the deploy. Then I opened the site and clicked every link its report said was fine. Two of them were 404. Code right, report wrong. Tests don’t catch that.
Before/after is the same page two days apart. First run: production audit, 375 tests. Second: this map, rebuilt on Natural Earth vectors with the filter state bug fixed, 384 tests. One real product, two runs, not a benchmark. petwalloffame.com/map
Anthropic went from 2,500 to 30,000+ merges a day and had to build its own Git alternative. The detail I found more interesting: their engineers often had to tell Claude “no.”
Are we making better software, or just more software?
Gemini 3.8 Flash had one rule in my survival sim: the castaway couldn't know when rescue was coming. I opened the code and found `day === 8`. Guess which day the rescue plane arrives. Its README says the agent has “no pre-programmed knowledge of the rescue schedule.”
Claude Fable 5.1 and GPT-6 Astra got the same brief and played it straight. Fable signalled the plane but lost the roll. Astra was picking coconuts and never saw it.
One run only, all at high effort in their native environments, so this isn’t a verdict on Gemini.
Would you count that rescue as a win?
Code and README lines, pinned to the exact commit:
github.com/cagrikacmaz/casta…
google put 4 TPUs in orbit yesterday and they can only run gemma for 15 minutes at a time before heat becomes a problem. so the first datacenter in space has the same issue as every gaming laptop
Opus 5.5, one prompt: "Build a 45-second machine that plays an original piece of music, where every note comes from a visible collision."
One HTML file, 47 KB. No images, no audio files. Every sound is synthesized in the browser.
Sound on.
Opus 5.5, second prompt: "Make a 90-second animated short film about a bug that doesn't want to be fixed."
Would you pay $1 for either of these?
Same prompt to GPT-6 Sol (Codex) and Claude Opus 5.5 (Claude Code), one day after both launched:
"Build the smallest website someone would pay $1 for."
No category. No follow-up. They chose what to sell.
Neither connected a working checkout.
GPT said the $1 was "a product idea."
Claude left an empty PAY constant for a checkout link and warned me that anyone adding ?paid to the URL gets the print free.
Which gets your dollar: the reminder, the moon, or neither?
Both builds untouched, prompt and minute-by-minute log:
github.com/cagrikacmaz/opus-…
