When the answer is not good enough
Fix the failure you can see. Do not automatically respond by making the system prompt longer.
The answer is generic
Add decision-relevant context, a specific audience, an example or the material the answer must use. Define what good means in this situation.
The answer is too long
Set priorities and a hard delivery shape: “Lead with the recommendation. No more than five points. Give the consequence and next action for each.” Remove repeated style rules.
The model invents facts or sources
Supply the sources, require a source for each claim, separate documented findings from inference, and allow “insufficient evidence”. Verify important claims outside the model.
The format keeps changing
Provide an explicit template or a few consistent examples. Use Structured Outputs in the API when another system will read the result.
The agent changes too much
Define what the request authorises, what is in and out of scope, which files or systems may change, and when approval is required.
One prompt has become enormous
Separate durable policy, the task brief, source material and user input. Split a long process into stages when intermediate results need review.
The model agrees with everything I suggest
Ask for an independent assessment against named criteria, evidence and plausible counterarguments. Avoid both “agree with me” and artificial “be brutal” instructions. Research on sycophancy shows that models can adapt to a user's stated view at the expense of truthfulness.
Long conversations keep getting more expensive
Measure input, cache reads, cache writes, output and tool calls per turn. Compact or start a new bounded workspace when old history no longer affects the task. Keep stable prefixes identical when using caching.