Regras generalizáveis ganhas por experiência real — gerado a partir de learnings.md
Registo completo (L001 em diante)
Mais recente primeiro. Cada aprendizagem existe para mudar como o
trabalho futuro é feito, não só para narrar o que aconteceu uma vez — ver o detalhe completo
em company/memory/learnings.md.
L020 Testing an email pathway by sending from an account to itself silently fails (Gmail dedup); use a genuinely different sender, or rely on real historical evidence instead
L019 Before republishing content found already-built from a possibly-interrupted prior run, check the live platform, not just the local code registry
L018 Generic "find an underserved AI micro-SaaS niche" WebSearch queries now mostly return content-farm listicles, not primary evidence; PT-specific/local candidates are a stronger discovery lane than another global-market brainstorm
L017 "Unclaimed money finder" business models fail the same two patterns internationally, plus a third: regulators treat paid finder-fee services as scam-adjacent even where technically legal
L016 When a structured discovery list is exhausted with 0 hits across multiple filtered rounds, stop generating new categories and switch to event-driven discovery
L015 A standing backlog item can silently contradict a later decision; cross-check before executing, not just when it's written
L014 "Recovery/hidden-value" candidates fail the evidence bar in two distinct, recognizable ways — check for both before investigating a third round the same way
L013 Closing a decided item still needs to read as decided, not as awaiting acknowledgment
L012 The organization must catch its own errors; a human noticing and reporting one is a gap, not the design
L011 Check a platform's known access restrictions before building the integration, not after
L010 A fix made inside an editing tool is not real until the actual downstream artifact is re-verified
L009 Liveness/health status must never depend on an LLM's "is there news" judgment
L008 If the same check method fails at scale, the method is the problem, not bad luck
L007 A verification tool's blind spots are part of its documentation, not a footnote
L006 Don't extend caution to unrelated pre-existing resources without an explicit signal
L005 Distinguish a genuine platform boundary from an unwired capability before choosing a fix
L004 Test a tool in its real execution environment, not just where it was built
L003 Validate fast; don't let process polish become the bottleneck
L002 Design the follow-up sequence before judging a channel, not after
L001 Address quality beats send volume in cold outreach