The architecture spike that listed rate limit strategy early prevented abusive traffic surprises in launch week. In AI startups (thread 10): We learned that writing 'rollback criteria' in migration plans reduces bridge thrash at night. The hardest bug lived between two services owned by two teams with two backlogs.

We learned that empathy without accountability still ships late. In AI startups (thread 10): The smallest improvement to date pickers reduced timezone bug reports from global users. The mentor who said 'tell me the worst case' before launch calmed the room usefully.

We measured the wrong thing first, then optimised ourselves into a corner. In AI startups (thread 10): We learned that customers trust contribution-based profiles more when they can see which circles shaped the signal honestly. We should have invested in backup verification jobs that restore to scratch weekly automatically.

We should have deleted unused invite links pointing at deprecated onboarding flows — confusion compounds quietly. In AI startups (thread 10): The clever abstraction blocked new hires for weeks; boring code shipped. The quiet win was aligning on a single moderation escalation path across time zones — fewer duplicate actions and fewer misses always.

The flaky dependency upgrade blocked releases until we pinned versions honestly. In AI startups (thread 10): We should have asked support what they hear before prioritising the roadmap. Design said edge case; support said thirty percent of tickets — words matter.

A single shared glossary reduced meetings more than any new dashboard. In AI startups (thread 10): The mentor who said 'write the decision log entry now' prevented repeated rehashing monthly. The best onboarding includes a guided first failure in a safe sandbox.

We learned that customers trust roadmaps that include maintenance and reliability work visibly. In AI startups (thread 10): The architecture diagram updated monthly beat the one updated once at kickoff. We stopped shipping 'temporary' SQL views that became analytics truth accidentally.

We learned that customers appreciate when CercleWork ships calm defaults for notifications instead of growth hacks noisy. In AI startups (thread 10): The flaky health check masked a partial outage — health checks need depth sometimes. We learned that small trustworthy releases beat big risky bangs for morale.

The smallest improvement to date formatting reduced international support confusion. In AI startups (thread 10): We should have invested in staging data refresh before the compliance audit panic. We learned that writing 'definition of done' with QA prevents last-minute thrash.

We learned that writing 'definition of done' with QA prevents last-minute thrash. In AI startups (thread 10): We learned that small wins for internal users compound into external velocity. We should have named a backup approver for production deploys before vacation season.

Reading old tickets was archaeology that paid better than guessing anew. In AI startups (thread 10): The migration plan assumed humans read email; they did not — multi-channel comms won. The quiet win was aligning on a single severity matrix across eng and support.

We learned that 'temporary' flags need owners and expiry dates in writing. In AI startups (thread 10): The quiet win was standardising environment names across repos and dashboards. We learned that naming owners for feature flags prevents zombie toggles nobody dares remove.

We stopped confusing 'busy' engineers with 'fully utilised' capacity for planning. In AI startups (thread 10): The smallest improvement to bulk export progress bars reduced 'is it stuck' tickets. Trust regrows slowly after a bad outage; over-communicate while it heals.

The integration that validated markdown sanitisation for replies prevented XSS surprises in public circles quietly always. In AI startups (thread 10): Trust regrows slowly after a bad outage; over-communicate while it heals. The quiet win was aligning on a single definition of 'active user' across teams finally.

The flaky deployment that ignored read replica lag taught us to surface replication delay in UI for sensitive actions quietly. In AI startups (thread 10): We stopped treating reliability work as invisible glue and started tracking it visibly. We should have invested in load testing the auth rate limiter before a viral post.