Masheed Ahadi (@masheed) header
Masheed Ahadi (@masheed) profile picture

PM that builds. Shipping AI agents into production systems.

SF Bay Area
Joined July 2011
Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)Photo by Masheed Ahadi (@masheed)
when i scope a task for my coding agent i list the files it's allowed to touch. if it needs one outside that list it stops and tells me why. a helpful cleanup two folders over is the part of a diff i'm least likely to read closely, so i'd rather it ask first
5
2
95
new model dropped this week and i'm not swapping it in yet. first i rerun a set of real cases from our product and only read the ones where the answer changed. those tell me way more than the benchmark chart bc they're my users' weird inputs
33
one rule i give my coding agent: when a test fails, tell me if you think the test is wrong or the code is wrong before touching either. left alone it'll sometimes just edit the assertion until it's green, and that's the worst kind of passing
1
1
81
before i read a diff from my coding agent i ask it one thing: what were you unsure about. it'll usually name a guess it made, like a field it assumed was optional. that's where i start reviewing instead of reading top to bottom.
2
3
162
Masheed Ahadi retweeted
marketing is 100x harder than coding
270
114
54
1,450
67,915
when i hand a bug to my coding agent i give it the repro steps and the exact wrong output, and leave my theory about the cause out. when i include my guess it tends to go fix what i suspected instead of what's actually broken.
2
89
Masheed Ahadi retweeted
Underrated life advice: Always assume things will work out, then do the work to make it true. It’s a mindset of active optimism. Earned faith. It creates a quiet confidence that allows you to tolerate the uncertainty of life better than anything else.
252
1,333
61
11,498
362,140
I believe AI shopping agents are about to split retail in two: platforms that let an agent complete a purchase, and walled gardens that don't. This basket owns the open side only, led by Shopify and Walmart, alongside marketplaces and storefront software courting agent traffic, a small merchant-data position, and a modest payments stake for the checkout button itself. Positions stay concentrated and no single name dominates, with cash held back to add if the rollout stumbles before it accelerates. supertake.com/t/r2o3wNMAF6U?…
1
2
129
small habit that's saved me a lot of digging: when an agent changes a record, log the old value next to the new one, not just "updated". when something looks off later, the first question is always what was it before, and that answer shouldn't need a replay.
35
One of the agents i'm building kept falling for this: a new email quotes an old thread and it looks like something just changed. the fix: split the quote out, keep its original date, and only treat the new lines as a change. been using @claudeai and @typesafeai's Jev. Do you strip quotes in code, or expect the model to just know?
51
Masheed Ahadi retweeted
230
3,152
178
27,297
625,073
For any agent task longer than a few seconds, I show the steps as they happen instead of a spinner. I think people wait much more patiently when they can see the work, and they often catch a wrong turn, like the wrong account picked in step two, before the whole run is spent.
43
When I change an existing screen, I give the coding agent the current version next to the new spec and point at what's different. Describing the whole screen again invites it to redesign parts I never asked about, and then I'm reviewing changes nobody wanted.
1
31
Before I build the happy path, I write the message a user sees when the agent can't finish. It says where it stopped, what it already changed, and what to do next. I think that screen holds most of the trust in an agent product, and it's usually the last one designed.
38
Masheed Ahadi retweeted
Every time we have a new model, I run the same simple prompt on all my projects: "Read this entire project and find everything that’s broken, inconsistent, inefficient, or no longer needed. Fix what you’re sure about and list the rest for me to review." highly recommend trying
352
539
229
10,086
1,181,483
When coding agents open more changes than I can read, the bottleneck is what I'll trust without a full pass. Mine is green tests on the risky paths, a hard deny for migrations and secrets, and a person for everything else.
1
59
When I reject an agent's change, I write one line on why before I move on. That list becomes the next eval set, because those are the cases where my judgment and the model already disagreed.
71
Small design habit for agent features: before building, I write the status lines the user sees while it works, like "checking your last three invoices". If I can't name a step in plain words, that step usually isn't defined well enough for the agent either.
63
When I scope an AI feature, the first thing I write is what it should say when it can't tell. A lot of the design work ends up there, because the model will answer confidently on inputs it shouldn't, and the real question is how the user notices and what they do next.
62
When my coding agent's second fix fails the same way, I stop prompting and reread my spec. By then the problem is usually something I left out, like an edge case or a rule that only lives in my head, and a third attempt just piles more code around the gap.
70