name: remote-llm
description: Craft prompts for Terry to run on local/work LLMs when code can't be shared directly. Use when Terry mentions proprietary code, work systems, or needs help with code he can't paste.

Remote LLM Workflow

Use this pattern when Terry needs help with code he can't share directly (e.g., proprietary bank code). Claude crafts prompts, Terry runs them on a local/work LLM (like Qwen3-32B at CITIC), and shares results back.

Workflow

1. Discovery prompt — Find the relevant code
2. Ask for file structure, class names, method signatures

3. Keep scope narrow to avoid overwhelming output

4. Terry runs locally — Pastes prompt + attaches files to work LLM

5. Interpret results — Understand what was found

6. Modification prompt — Make specific changes

7. Reference exact locations from discovery

8. Keep changes incremental

9. Iterate — Repeat until task complete

Prompt Delivery

Create secret gists for easy copy-paste:

gh gist create -d "discovery prompt" -f "prompt.md" - <<'EOF'
[prompt content here]
EOF

Always secret gists. Never -p or --public. Delete after use.

Prompt Structure Tips

• Start with clear context: "You are helping modify [system/module]..."
• Be explicit about output format: "Return only the modified function"
• For discovery: "List all methods in [class] with their signatures"
• For modification: "In [file], find [function], change [X] to [Y]"

Model-Specific Notes

If Terry mentions which model, check vault for tips:
- Qwen3: See [[Qwen3 Prompting Best Practices]] — thinking mode, temperature settings