We should have deleted unused IAM policies quarterly — stale permissions accumulate quietly. In ML practice (thread 18): The bug bash found issues our automated suite could not imagine — humans still matter. The architecture spike that listed multi-cloud egress costs prevented surprise bills honestly.

We underestimated how long humans take to trust a new workflow. In ML practice (thread 18): The best engineers I know admit 'I do not know' quickly and learn faster. We learned that humour about meetings lands better when paired with a concrete experiment proposal.

The build cache sped CI until it served stale artifacts — trust but verify. In ML practice (thread 18): The architecture decision to prefer idempotent handlers aged better than 'exactly-once' dreams. The smallest UX copy change cut support tickets more than a new API field.

We should have invested in offline-friendly read modes before pitching global teams with unreliable connectivity honestly quarterly. In ML practice (thread 18): We stopped shipping 'just a quick script' without code review because scripts run in prod too. Remote made async communication non-optional; emoji tone-misreads were real incidents.

We finally aligned on a single severity rubric and stopped arguing adjectives. In ML practice (thread 18): We learned that small improvements to internal wikis reduce repeated onboarding questions weekly. The quiet win was aligning on a single definition of 'active user' across teams finally.

The mentor who said 'prove funnel conversion with experiments' sharpened growth marketing debates usefully. In ML practice (thread 18): Cut scope early; nobody remembers the sixth nice-to-have you skipped. Good incident comms reduce duplicate tickets more than faster fixes sometimes.

The mentor who said 'write the customer comms before you merge' improved launch discipline. In ML practice (thread 18): We should have asked data science about seasonality before promising growth curves. The mentor who said 'show the customer quote' ended abstract prioritisation debates.

The integration that validated markdown sanitisation for replies prevented XSS surprises in public circles quietly always. In ML practice (thread 18): We learned that sustainable pace is a feature, not a luxury for later. The architecture review that asked about multi-tenant isolation assumptions caught a leak path.

The integration that surfaced partial failures prevented silent half-updates in billing. In ML practice (thread 18): The architecture review that asked about backup restores caught a real gap. The design that included offline states first saved rural users real frustration.

We stopped shipping 'temporary' IP allowlists that became permanent security theatre. In ML practice (thread 18): The best teams treat documentation updates as part of definition of done for features. We merged on Friday once and the meme became policy faster than any memo.

We learned that gratitude in tickets is cheap and improves cross-team goodwill. In ML practice (thread 18): We should have invested in load testing the auth rate limiter before a viral post. Rubber-stamping reviews to be nice is not kindness to the person on-call.

The flaky deployment that ignored canary latency regressions taught us to watch p99 not only errors. In ML practice (thread 18): We underestimated how much cognitive load a second deployment pipeline adds. The smallest improvement to date pickers reduced timezone bug reports from global users.

We stopped confusing 'more circles' with 'healthier network' when measuring product success honestly quarterly. In ML practice (thread 18): The mentor who said 'show me the reply latency distribution' grounded reliability debates for discussion products helpfully. Half the team knew the risk; nobody felt authorised to say stop on the call.

We learned that writing 'non-goals' in RFCs prevents zombie scope resurrection. In ML practice (thread 18): The architecture review that asked about multi-region assumptions caught naive defaults. The smallest improvement to CSV export headers reduced analyst rework weekly.

The smallest improvement to moderator bulk actions reduced time-to-clean spam bursts measurably during attack weekends always. In ML practice (thread 18): The clever abstraction blocked new hires for weeks; boring code shipped. The architecture spike that listed compliance constraints early saved redesign pain later.