search-atlas-group cli-llm-routing
Prefer installed CLI tools for ad hoc AI consultation. They use your normal login flow, keep credentials out of prompts, and make the command visible.
Use A Second Model When
- the task involves current documentation;
- you need a skeptical architecture review;
- the context is too long for the current session;
- the answer affects a large amount of work.
Do Not
- paste API keys into prompts;
- call provider REST APIs for one-off consultation;
- run shell commands suggested by another model without reviewing them;
- assume the second model has your full chat context.
Prompt Shape
Context:
<short, sufficient context>
Question:
<specific decision or review request>
Return:
<bullets, risks, recommendation, sources if web was used>
Output Handling
Treat the result as advice, not authority. Pull useful reasoning back into your main session and verify commands before running them.