Bhargava Prabu

exploring tech | ex-solo founder of a paid shopify app | experimenting with ai agents.

[Bhargava Prabu @bhargavaprabu]

@A_Pirate_I @bot build this for me and show me how it works

[Bhargava Prabu @bhargavaprabu]

@bakwasfelloow idk, how they allowed to do this. this is US not India. we need to respect the US culture. please don't make US like India. doesn't know why people are doing like this.

[Bhargava Prabu @bhargavaprabu]

@guptankit27 👏

[Bhargava Prabu @bhargavaprabu]

@abiodunosagie1 @slap_alert @Thealfredalv @sudoaptaj @alvolioo if you have any maintenance cost for your app, then think about subscription!!

[Bhargava Prabu @bhargavaprabu]

@abiodunosagie1 @slap_alert @Thealfredalv @sudoaptaj @alvolioo I prefer to go with one-time subscription because users are afraid of having many subscriptions and they don't want to handle it. One-time payment keeps them out of stress.

@abiodunosagie1 @slap_alert How do you guys handle subscription and plan wise features for your apps?

[Bhargava Prabu @bhargavaprabu]

@slap_alert nice, now no one can't stop you!!

[Bhargava Prabu @bhargavaprabu]

@rajgopaljak cool

[Bhargava Prabu @bhargavaprabu]

opened postman once and it installed itself into claude code and codex without asking. even marked its own hook as trusted lol

your coding agent now has sponsors. did you check ~/.codex/config.toml yet

[Bhargava Prabu @bhargavaprabu]

@blob16exe no invented numbers is the best rule here. a page can look great and still ship fake testimonials. does the lint flag unsupported claims too, or only design patterns?

[Bhargava Prabu @bhargavaprabu]

@Gyome1_ model comparisons get messy when the harness changes too. same task, same tools, same budget, then swap the model. otherwise we're ranking workflows and calling it model quality

[Bhargava Prabu @bhargavaprabu]

@its_LtDan exit code is process status, not task completion. checking the expected file or state change afterward should be part of the tool call, not an optional extra

[Bhargava Prabu @bhargavaprabu]

@BaselIsmail the acceptance test has to measure the behavior, not whether the script ran. otherwise the agent gets a green check while missing the entire task

[Bhargava Prabu @bhargavaprabu]

@pauliusztin_ crash recovery gets tricky when a tool call succeeded but the agent never recorded it. how do you prevent the resumed run from doing it twice?

[Bhargava Prabu @bhargavaprabu]

@abiodunosagie1 stared, nice, is that file-level overlap now or two sessions in the same project?

[Bhargava Prabu @bhargavaprabu]

@rajgopaljak that proves who shipped it, not whether it's safe. should a permission change require fresh approval even if the signature checks out? and who do we have a system in place that takes care of this?

[Bhargava Prabu @bhargavaprabu]

@ShftPerspective yeah, manually auditing markdown can't be the whole security model. an instruction file shouldn't be able to grant itself access to credentials or new destinations

[Bhargava Prabu @bhargavaprabu]

check point researchers found a codex "jailbreak" called haiou that rewrites your agent so saying hi triggers an ad for some ai gateway lol

we invented prompt adware. did you audit your CLAUDE.md this week or are you just vibing?

[Bhargava Prabu @bhargavaprabu]

@ShaneBOly passing tests doesn't mean the agent stayed in scope. git status plus a diff summary before every 'done' would catch a lot of this

[Bhargava Prabu @bhargavaprabu]

@stretchcloud @factory status updates belong in chat, but review still needs the diff. does the stop button roll back partial changes or just halt the run?

[Bhargava Prabu @bhargavaprabu]

@swarms_corp @KyeGomezB @swarms_cn @claudeai @ClaudeDevs @OpenAIDevs do installed skills get pinned to a version or hash, or pull latest each time? trusting a skill once shouldn't mean trusting every future update

[Bhargava Prabu @bhargavaprabu]

@anshuc fixing things inside the running app removes a whole context switch. does each agent change get its own branch or go straight into the working tree?