𝐙𝐨𝐡𝐚𝐢𝐛 𝐀.

designing and building brand, product and ai experiences. commerce for games at @neoncommerce. vr f1 racer.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@vercel_dev FTP❤️

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@viktor_pali but will be cool to have the subscription based agents to interface with the docs, i.e. not pay any additional api fee per usage

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@thericebowlgirl yess

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@ariccio @skeptrune using their own apps like codex and claude code are fair use so tou can symlink your vault as the operating layer with your portable skills and knowledge systems. be agent independent

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@verymidengineer codex settings -> profile

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

codex says i’ve used 7.4bn tokens. every api pricing calculation is making it between $70k - $400k . the model was gpt 5.5 pro.

no wonder companies are running out of their ai budgets. it makes no sense for using it outside subscriptions.

enjoy the vc subsidized tokens.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@karrisaarinen the ones showing what they built aren’t talking about building with ai even if they are. because they know it’s about the product not the tool- and that tool actually makes the product seem less trustworthy

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@varadh https://t.co/61S7wrOKJr

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

not sure what category this fits in. i’m running research agents which use skills on how to research and how to deploy via inhouse mcp. and these skills are kind of self improving.

layer 1 agent can update the codebase/mcp/services/api’s

layer 2 is the research orchestrator who starts research tasks per agent giving them skills and targets

layer 3 is the research agent

—

layer 3 does the research, use the mcp to push that research and report back to layer 2 if they are blocked or their process if they’re done

layer 2 often unblocks that layer 3 agent as they have more qa oriented skill rules available to them, only after a designated batch recommend to update the research skills. can escalate to layer 1 agent as well. layer 2 spawns new agents as well with a strict rule of minimal repeatable initiation promps. without extra inline instructions so the process remains scalable.

layer 1 agent gives recommendations if mcp guide or skills need to be updated as well or add deterministic validators in the codebase.

i choose to be in the loop for recommendations that are not obvious bugs but more like behavior change of the skills. this ensures the skills don’t become a dump for one-off edge cases.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

not sure what category this fits in. i’m running research agents which use skills on how to research and how to deploy via inhouse mcp. and these skills are kind of self improving.

layer 1 agent can update the codebase/mcp/services/api’s

layer 2 is the research orchestrator who starts research tasks per agent giving them skills and targets

layer 3 is the research agent

—

layer 3 does the research, use the mcp to push that research and report back to layer 2 if they are blocked or their process if they’re done

layer 2 often unblocks that layer 3 agent as they have more qa oriented skill rules available to them, only after a designated batch recommend to update the research skills. can escalate to layer 1 agent as well. layer 2 spawns new agents as well with a strict rule of minimal repeatable initiation promps. without extra inline instructions so the process remains scalable.

layer 1 agent gives recommendations if mcp guide or skills need to be updated as well or add deterministic validators in the codebase.

i choose to be in the loop for recommendations that are not obvious bugs but more like behavior change of the skills. this ensures the skills don’t become a dump for one-off edge cases.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

we’ve been actively trying to automate parts of design that we didn’t want to do like slide decks, newsletters, one pagers and set up the design systems in claude design.

before we did this- everyone was saying design is becoming a bottleneck. make us independent. now despite having all of these tools they still come back to us.

turns out we designers have been underselling our value so much. it was never just about the make things pretty stuff we did. it was about the extra strategy work we did for free that execs in other domains charge 4-10x what they pay designers for.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

my new goal is to figure out within the next 12 months to self host a full kimi level model and afford it.

the subscription based tokens looks like its days are numbered.

these models are getting so deeply integrated to the workflow that soon enough we’ll be held hostage and have to pay more for the computer than our monthly salaries.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@tonyseets great holistic overview of what it’s like to get started with building with AI 👏

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@aatif_awan this is so good! great work.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@jturntdev definitely better than claude’s. we were just spoiled

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@Prathkum and you put everyone on a lower tier account, whereas the ones who need it need the max plan.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@jacobmrt @galluzzo_julian they don’t not care about good design. they just don’t have patience or imagination for it. they have to experience it.

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

in my last two projects i’ve borderline offended people, broken out of the briefs. and that’s why they ended up being successful.

i knew what they thought they wanted wasn’t good for them and would have kept us in revision hell.

you can’t prompt this stuff. it took me 20 years to get here

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@forgebitz they still exist and continue to ruin the world

[𝐙𝐨𝐡𝐚𝐢𝐛 𝐀. @HeyZohaib]

@faizansiddiqi i’m saying the IT firms are also mostly selling repackaged with them in the middle remote jobs only.

so even their definition of calling themselves IT firms to remove incentive from remote job folks is wrong.