.tips
page 3 of 11, 261 small things in all. back to the first.
.august
2026, 25 tips.
Every agent run has a turn limit and a token budget
ai-agentsreliabilityA test that fails for more than a day gets fixed or deleted
testingengineering-practicePrompts live in the repository with a version and a changelog line
ai-engineeringengineering-practiceKeep a map file the agent reads every session
ai-agentsmaintenanceTemperature 0 for anything with a right answer
ai-engineeringllmFew-shot examples come from the eval set, not from imagination
ai-engineeringllmRead the code for ten minutes before the first prompt
ai-agentsengineering-practiceA system prompt is a job description, not a manual
ai-engineeringllmBenchmark a runtime with weighted production traffic
nodejsperformancebreakpoint() drops you into the debugger where the bug is
pythondebuggingasyncio.TaskGroup, not gather, for a group of tasks
pythonconcurrencyCommit messages do not need the frontier model
llmlocal-modelsfunctools.cache on pure functions that are called repeatedly
pythonperformancePrefix every chunk with its document title and heading path
ragsearchtyping.Protocol for 'anything with these methods'
pythonlogging, not print, from the first line of a service
pythonobservabilityMeasure retrieval on its own before touching the model
ragtestingpathlib, not os.path string surgery
pythondataclass(slots=True, frozen=True) for value objects
pythonVector search loses on names. Add keywords
ragpostgresqlsearchOne behaviour per test, however many assertions it takes
testingSeven days for successes, ninety for failures
observabilitycostRead coverage as a map of missing tests, not as a score
testingengineering-practiceRun the tests that touch what you changed, then everything in CI
testingdeveloper-experienceSample traces at the tail, keep every error
observabilityopentelemetry