xuefeng-method

Replace exhaustive case enumeration with behavioral clustering: group user actions into 5-15 typical patterns, design an AI response strategy per cluster, and add graceful degradation for edge cases. Ship at 80% quality and iterate using real user feedback.

Assign each agent exactly one responsibility, such as understanding, execution, verification, or expression. Split when functions are orthogonal with independent validation criteria; avoid splitting when there are context dependency chains or complex merge alignment requirements.

Monitor four signals: output distribution shift versus baseline, declining user satisfaction, rising anomaly rates like timeouts and hallucinations, and inter-agent inconsistency. Compare current output statistics against a first-week baseline and alert when drift scores exceed thresholds.

Use xuefeng-method when user input is open-ended, AI drives core decisions, and outputs are probabilistic. Use keqian-method when scenarios are well-defined with enumerable input-output combinations. Hybrid products can apply both per module.

Open user behavior multiplied by probabilistic model outputs and dynamic context creates combinatorial explosion that cannot be enumerated. Use property-based checks like verifying output belongs to an acceptable set instead of asserting exact equality.

Route by task type: codex models for pure coding, general flagship models for mixed reasoning, and lightweight models for simple execution to control cost. Compact context before hitting dumb-zone thresholds, around 0.5 of the window for domestic models and 0.7 for top international models.