billzh

building @tribedotrun (@AllianceDAO ALL13)

[billzh @billzh]

building a new social app on Solana and I'd love your help if you are a degen & into socialfi

crypto has long promised us the ability to bet on underrated chads and get rich. any account's worthiness should be measured in market price, not Twitter followers

but today this is still an unsolved problem. most prior attempts were too extractive and executed poorly

I don't have all the answers but I'm experimenting and plan to iterate fast, starting with small degen chats

so if you are interested in trying out the alpha version & giving feedback, please dm

[billzh @billzh]

@QwQiao what prompts do you use to generate thesis vs antithesis?

[billzh @billzh]

@bennett2b is there a Rust sdk??

[billzh @billzh]

@kr0der Devin review

[billzh @billzh]

@thsottiaux please help - codex is killing my mac

[billzh @billzh]

@_lopopolo is the lever mostly the prompts in the skill? or it can add cli / scripts too

[billzh @billzh]

I dump most of my mid / small backlog items directly to my agent and just review "proof-of-work"

no need to manually manage claude codes (which is now reserved for hard tasks that require a lot of thinking / planning)

[billzh @billzh]

I realized that I need an AI engineering manager instead of coding agents

then I can just dump my backlog on Telegram and ask him to baby sit all prs

[billzh @billzh]

https://t.co/nQS5Klr6QW

[billzh @billzh]

built an agent internally called Mando

it auto manages a swarm of claude codes to work through my entire backlog - babysitting each one until it produces e2e verification proving the work is done

it monitors progress, nudges claude when it's off track and always enforces quality gate

alI I need to do is review the final pr and send feedback on Telegram if it's not 100% there

runs locally on my Mac. basically an orchestration layer that gives me a tech lead that never sleeps

I get moved up the stack - from managing code to managing tasks

try it out. github link below

[billzh @billzh]

how do you make codex talk more naturally?

its response has too many long-winded slops that no one should ever read

[billzh @billzh]

I realized that I need an AI engineering manager instead of coding agents

then I can just dump my backlog on Telegram and ask him to baby sit all prs

[billzh @billzh]

so who else is organizing their sleep schedule around claude peak hours??

[Thariq @trq212]

To manage growing demand for Claude we're adjusting our 5 hour session limits for free/Pro/Max subs during peak hours. Your weekly limits remain unchanged.

During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before.

[billzh @billzh]

recently I just ask my claude to go autopilot after we have created a plan

let's go. go into autopilot mode and implement everything, verify, and iterate. make sure you have implemented every item in [plan file] and custom e2e completed verification without any issues. as a last step, run /x-pr and address all PR comments.

this basically allows me to hand off the task to claude and come back in 1 hour to check verification results (recoding or screenshots) in pr

[billzh @billzh]

@code_cowboy https://t.co/CU4f5B4OoP

[billzh @billzh]

apparently claude code has a tendency to avoid doing work by claiming something is out of scope

so I added a stop hook that detects keywords like "pre-existing", "defer", "separate issue" etc

it nudges the model into fixing everything it finds (but doesn't block the thread)

[billzh @billzh]

I'm glad that we never need to touch figma again

just run 8 agents using @paper w/ Impeccable's claude skills

[billzh @billzh]

claude code has another two annoying habits that I guard against

- useless backward compat / legacy code for no reason

- answers w/o runtime evidence (basically hallucinating), esp. when it uses the word "probably"

stop hooks here are more effective than claudemd rules

[billzh @billzh]

apparently claude code has a tendency to avoid doing work by claiming something is out of scope

so I added a stop hook that detects keywords like "pre-existing", "defer", "separate issue" etc

it nudges the model into fixing everything it finds (but doesn't block the thread)

[billzh @billzh]

@bcherny why is claude code taking ~500 GB with 5 tabs on iTerm2??

[billzh @billzh]

@ScottWu46 overall works great, but it seems to be really slow on larger PRs

[billzh @billzh]

@bcherny how many MCPs do you usually load into your context window? they occupy a lot of tokens

[billzh @billzh]

why is there a big slippage when bridging $hype from Solana on Wormhole??