Default Model Failure Signals

Default model failure signals are the moments when a team should stop pushing the same AI route and switch tools or verification mode. They prevent sunk-cost prompting, where the user keeps rephrasing a task that the current model is structurally bad at.

Common signals: repeated fabricated file names, inability to follow repository evidence, circular debugging suggestions, missing source citations for current facts, slow handling of bulk mechanical work, ignoring constraints after correction, or producing code that cannot be verified locally.

The response should not always be “use a stronger model.” Sometimes the right escalation is deterministic tooling, search, tests, a browser check, a smaller cheap model for bulk labels, or human review. The failure signal only says the current route is no longer sufficient.

Record the signal in plain language. For example: default failed because it could not inspect logs; switched to coding agent and ran tests. Or default gave unsourced API behavior; switched to official docs source check.

This habit improves model selection without requiring model fandom. The team learns which failures belong to the task, which belong to the tool, and which belong to unclear instructions.

Default Model Failure Signals

// COMMENTS

ON THIS PAGE