thingslist @thingslist

I ship Web, iOS, Android apps and games solo with AI. Bad at marketing, so I'm building an AI that does it for me. Both in public. San Francisco, CA Joined June 2013

Tweets

359
Followers

47
Following

421
Likes

913

thingslist @thingslist

3 hours ago

@Eli5defi Yeah I just keep one CLAUDE.md and paste it into whatever tool I'm in. Ugly but it works

0 0 0 7 0

View Details

thingslist @thingslist

3 hours ago

Agents that rewrite their own instruction files after a failure never check whether the rewrite helped. New paper, SkillProx.

0 0 0 1 0

View Details

thingslist @thingslist

2 days ago

@QuinnyPig It's the thing executing on my machine all day but ok

0 0 0 20 0

View Details

thingslist @thingslist

3 days ago

@bcherny ~0 on unseen attacks is a wild number. What did it cost in false positives on normal edits?

0 0 0 61 0

View Details

thingslist @thingslist

3 days ago

@SaraDiscovers Claude Code. I ship iOS apps solo with it, never had a reason to switch

0 0 0 47 0

View Details

thingslist @thingslist

3 days ago

@MacRumors Wait the sessions can talk to each other now? I've been running mine in separate terminals like a caveman

0 0 0 581 0

View Details

thingslist @thingslist

3 days ago

24 hours with Prime Agent, a new coding agent. Two lines apart: more efficient at context, twice the tokens.

I've been trying Prime Agent for 24h, here are my thoughts: 1. Having a single IPython tool is genius. 2. The agent is way more efficient at managing context, but it uses twice as many tokens. 3. The continual harness can over-fit to certain tasks, which is bad if you work on

43 13 456 34K 188

0 0 0 37 0

View Details

thingslist @thingslist

3 days ago

Someone in the DeepSeek V4 Flash thread: every provider except DeepSeek charges at least 10x for cached input, the repeat part of your prompt the model has already seen. In coding agents that's 90% of the tokens. So the cache rate is basically your whole bill.

0 0 0 38 0

View Details

thingslist @thingslist

3 days ago

@Sourcegraph Whether it sticks is the only number I care about. What was the revert rate?

1 0 0 72 0

View Details

thingslist @thingslist

4 days ago

@xn1cklas Careful with the innovative parts, anything non standard in my iOS apps needs a tooltip or people bounce

1 0 1 35 0

View Details

thingslist @thingslist

4 days ago

@SouthernValue95 Feels right from the inside. My usage went way up this summer and I didn't plan it, it just got good enough to leave running.

0 0 0 48 0

View Details

thingslist @thingslist

4 days ago

@chazm_ball Nah a 10 year old codebase does that on its own

0 0 1 3K 0

View Details

thingslist @thingslist

5 days ago

@DataChaz Zero moat maybe but I still pay for both. Nobody cancels, they just add another one.

1 0 1 266 0

View Details

thingslist @thingslist

6 days ago

80B parameters in 4.3 GB of RAM on a Mac. The 35B one runs on an iPhone.

0 0 1 16 0

View Details

thingslist @thingslist

6 days ago

@mynameisyahia Is it actually fast though? Arc got so slow for me by the end.

1 0 1 607 0

View Details

thingslist @thingslist

6 days ago

@GithubProjects So the harness is most of the gap? Kind of wild if a cheap model closes that much.

0 0 1 935 0

View Details

thingslist @thingslist

6 days ago

@commonsenseplay @GenZinvestor10 Like fsd.. lol

0 0 1 20 0

View Details

thingslist @thingslist

a week ago

Four things trending on GitHub today ship as agent skills instead of packages. A security tool router, a topic research skill, a Korean skill pack, a UI taste skill.