Abdul Builds

AI you didn't know you could do, in plain English. You know your business and I know the AI tools, so I teach you the ones that fit your work. Montreal, EN/FR.

[Abdul Builds @buildswithabdul]

companies made a lot of money from people forgetting to cancel.

now AI cancels for you. meta's muse finds every subscription on your card and kills the ones you forgot. rocket money's rowan texts you, you reply "cancel", done.

the catch: you're handing an AI your bank.

[Abdul Builds @buildswithabdul]

@arjunomics @kenbwork that's a fair split and a better question than mine. if the setup mirrors how an oncology office already works, the baseline is the office, and the eval asks whether agents hold up inside that workflow. a single-agent number would still show what the mirroring costs.

[Abdul Builds @buildswithabdul]

@sarymismail that answers it, and it's better than a reset button. a timeline of complete self-contained states means you can go back to the moment before a bad patch, scene included. curious how big those files get after a long session of patches.

[Abdul Builds @buildswithabdul]

@mark_k it does, and that's the new trap. the agent that wrote the code also writes the tests, and tests written to pass do pass. split them: one session writes tests from the spec before any code exists, another writes the code. the suite stops being a mirror.

[Abdul Builds @buildswithabdul]

@eugeniu_ghelbur history section with a date is exactly the trace. one more field worth adding: which agent wrote the entry. when eight of them share a vault, knowing who changed the fact is how you find the one with the bad source.

[Abdul Builds @buildswithabdul]

@robertoblake fair, i meant expensive in effort. someone had to sit with a hundred talking heads and write down where the first cut lands. you did that part, which is why a free bundle of yours beats a paid mcp with a blank prompt.

[Abdul Builds @buildswithabdul]

@tricalt docstrings as first-class graph docs is the right call, with one trap: docstrings lie. they describe the function as it was three refactors ago and the agent trusts the prose over the body. store the commit age of each next to its symbol so a stale one gets flagged, not followed.

[Abdul Builds @buildswithabdul]

@Imran_Razaq @krunalbuilds gateway limits protect the servers. they dont protect the one expensive route. put a global limit in front of an llm endpoint and one key hammering it starves everyone else while staying under the cap. per key, per route, where the cost is, then the gateway on top.

[Abdul Builds @buildswithabdul]

@0xKense1 the never lines are the ones to move out of the file. anything that must hold every single time belongs in a hook that runs outside the model, and the md keeps the preferences. do that and the file gets short on its own, because most of its length was pleading.

[Abdul Builds @buildswithabdul]

@kevinwhinnery @OpenAIDevs the clicks were never the hard part. the session is. a fresh profile hits a login wall on step two and a human babysits every run. curious whether the sdk can attach to an existing logged-in browser over cdp instead of spawning a clean one.

[Abdul Builds @buildswithabdul]

@SimonasLTU1 the leftover goes furthest on things you would never spend a prompt on: a subagent reading the whole repo for dead code, tests for the files nobody touched, a review pass on last week's merges. the limit resets, the test suite stays.

[Abdul Builds @buildswithabdul]

@ziv_ravid the code on a fresh file reads about the same as last year. what changed is the reroll rate: fewer tasks come back needing a second prompt, and that compounds more than code quality does. count accepted diffs per task, that is the number that actually moved.

[Abdul Builds @buildswithabdul]

@PovilasKorop the example that converts is the one that looks like their codebase, and docs almost never have it. what sold me on the last tools i paid for: a copy-paste snippet per framework on the homepage, real output under it. see the result before installing and pricing is a formality.

[Abdul Builds @buildswithabdul]

@defileo the 2000 freeware is the easy case: no packer, no obfuscation, static imports, so the decompiler hands claude a clean call graph. run it on anything packed after 2015 and 30 minutes turns into a day of guessed function names. the win is real, the sample is soft.

[Abdul Builds @buildswithabdul]

@zachlloydtweets the small loops close because something red turns green. "what to build next" has no red. the closest stand-in is the support inbox and the session recordings, so the agent that reads those before it suggests anything is the one that closes the loop.

[Abdul Builds @buildswithabdul]

@rohanpaul_ai the accreditation is the moat and also the liability. a browser agent that books the wrong dates costs you a refund fight. an accredited agent eats the chargeback and loses rate access if it happens too often. so the hard engineering here is the confirm step before booking.

[Abdul Builds @buildswithabdul]

@realNatalieFort step 6 is the whole method. the tell that claude is guessing instead of tracing: it names a function that does not exist in the binary. make it paste the real symbol or address next to every claim, then grep for it. one missing symbol and the whole path above it is fiction.

[Abdul Builds @buildswithabdul]

@mark_k slop comes from merging the first draft. ai code that had to pass a test suite it did not write reads cleaner than most hand-written diffs. the sloppy dev next year is the one who merges without reading, whichever side wrote it.

[Abdul Builds @buildswithabdul]

@DagmawiBabi point 1 is where the fix already landed. anthropic's code execution with mcp post: expose tools as files the agent imports and writes code against, not a wall of schemas. their example cut 150k tokens of tool defs to 2k. the agent crafting its own tool is just that script.