Kushagra Mathur

Software Engineer | AI Builder
Turning ideas into products with code.
Sharing code, system design & career insights

[Kushagra Mathur @CoderGhost37]

Opus 5.5 week made one thing obvious: writing code isn't the scarce skill anymore.

Deciding what to build, writing a spec an agent can't misread, and knowing which tradeoffs matter at scale is the job now.

[Kushagra Mathur @CoderGhost37]

@theo At this point the ceiling usually isn't the model, it's how clearly you can describe what you actually want built. Spec quality is the new bottleneck

[Kushagra Mathur @CoderGhost37]

@theo Option 3: Microsoft never really ports the debt and just ships a translation layer good enough that nobody notices. That's what made Apple's switch work, not the porting

[Kushagra Mathur @CoderGhost37]

@theo Design cleanup is the most underrated performance upgrade. Same features, less friction, and the app suddenly feels faster

[Kushagra Mathur @CoderGhost37]

@Suryanshti777 Both are true because writing code was never the bottleneck. Vibe-coded apps don't break on syntax, they break on a fuzzy spec and a data model nobody thought through

[Kushagra Mathur @CoderGhost37]

@pauliusztin_ Harness over model is the most underrated take right now. With the same model, better context management and tool design move output up a whole tier

[Kushagra Mathur @CoderGhost37]

@arpit_bhayani Fine by me. Typing code was never the bottleneck.
Deciding what to build, what to cut and which tradeoffs you can live with in prod still is

[Kushagra Mathur @CoderGhost37]

@theo Speed is the most underrated feature in dev tools. Every extra second of waiting is an invitation to switch tabs and lose the thread

[Kushagra Mathur @CoderGhost37]

@mehulmpt The routing layer is the easy half. The hard half is keeping prompts and output formats portable, so failing over doesn't quietly change behavior

[Kushagra Mathur @CoderGhost37]

@sakatayasha Finish time matters less than how many follow-up prompts it took to get the puzzle working.
One-shot comparisons hide the iteration cost.

[Kushagra Mathur @CoderGhost37]

@bindureddy Cheap and fast on easy tasks is where most production traffic actually lives. The real design move is routing: send the boring 80% to Flash and only escalate to a frontier model when the task needs it

[Kushagra Mathur @CoderGhost37]

@levelsio The underrated part is running the same audit prompt across every box, since consistency beats a one-off deep dive. Next step is putting it on a weekly cron so drift gets caught before someone else finds it

[Kushagra Mathur @CoderGhost37]

@nikitabier The scary part isn't the voice, it's that banks still treat caller ID and "confirm your OTP" as identity. Calling back on a known number should be the default, not a tip buried in a PSA

[Kushagra Mathur @CoderGhost37]

@piyushgarg_dev The part people miss is that memory is really a retrieval problem.
Deciding what to save and when to inject it back into context matters more than the model, and bad memory is worse than none

[Kushagra Mathur @CoderGhost37]

@theo Multi-model review only works if one model's confident nonsense can't outvote the others.
Do you weight the reviewers or just flag where they disagree?

[Kushagra Mathur @CoderGhost37]

@striver_79 The rewrite always looks clean because it hasn't met production yet.
Those ugly if-statements are years of bug reports nobody wrote down

[Kushagra Mathur @CoderGhost37]

@striver_79 This is way closer to the actual job than LeetCode. Half the skill is knowing when the AI's fix is wrong, and you can't do that without knowing the algorithm

[Kushagra Mathur @CoderGhost37]

@Sumits_X I'd put idempotency and retries above microservices on this list.
A lot of prod incidents turn out to be a retry storm or a duplicate write, not a scaling problem

[Kushagra Mathur @CoderGhost37]

@rileybrown Every "is the new model dumber" thread is half real regression and half us getting used to it. A small personal eval set is the only way to trust the vibe

[Kushagra Mathur @CoderGhost37]

@theo Six months and it already feels like a different era. The hard part now isn't picking a model, it's keeping your evals and prompts from rotting every quarter