Abe K

Berkeley EECS / Founder of Pneuma (using AI + human data to product test for a fraction of the cost)

[Abe K @akarik3873]

NFC chips aura farm so hard when people don’t know what they are.

[Abe K @akarik3873]

@denizm_dev I haven’t noticed much but I typically try to clear around 500k tokens. In theory, Claude reinjects the Claude.md unless it’s over ~5k tokens. If ur compressing twice u probably still have a significant amount of context causing bloat.

[Abe K @akarik3873]

I struggled a lot with Claude’s verbosity for a while. Tried to tun down output style, tell it in chat, put it in the Claude.md. Finally realized it needs concrete actions. In my user level Claude.md: Response style: lead with the outcome, in at most 5 lines.

[Abe K @akarik3873]

Think I saw something similar somewhere, but forget where

[Abe K @akarik3873]

Tip for agents swarms: use cloud run jobs to run the agents so you don't get timed out by ur backend being too small (services limit is 60 min). Also helps to split up the swarm across containers cuz you can’t split across services. And u only need to pay what you use.

[Abe K @akarik3873]

- No 2 brick/stone next to each other
- No 6/8 on the same resource
- No two of same number/resource
- No 3 wheat sheep or wood next to each other
- No resource next to its port
- No of the same number next to each other
- No 6/8 next to each other

[Abe K @akarik3873]

Catan getting really heated so I made a board randomizer that makes the generation a little fairer. Hand coded most of it a while back.
https://t.co/D4mfZHULYt