LLM Prompting & Learning Guide

Friday. July 18, 2025 - 3 mins

LLM Prompting & Learning Guide

I spent today downloading and processing all of my chats with LLMs, with the purpose of compressing my previous prompts and using them to get an LLM prompt-review! I got some useful tips which I think might prove helpful for other researchers, then I’m going ahead and putting up this practical companion for improving efficiency, learning, and collaboration with language models.

---

Strengths in LLM Interactions

1. Precision in Technical Queries

• Clear and concise code-related queries.
• Specific task-driven requests (e.g., “make this function a one-liner”).

2. Depth of Exploration

• Engage conceptual boundaries (e.g., AI agency).
• Stress-test LLM capabilities (e.g., “solve Riemann’s hypothesis”).

3. Iterative Refinement

• Follow up to clarify or correct model outputs.
• Request simplifications, optimizations, or alternative approaches.

4. Multimodal Use

• Apply LLMs across disciplines: code, music, philosophy, bureaucracy, etc.

---

⚙Areas for Improvement

1. Add Context to Abstract Queries

• Add background.
  Example:

  “In Transformer architectures, why must Q/K matrices share dimensionality? Explain with attention-score normalization.”

2. Avoid Over-Fragmentation

• Batch related ideas.
  Example:

  “Implement | and |= for my AronsonSet class. Ensure iter_dict updates.”

3. Clarify Ownership of Requests

• State whether you want critique, explanation, or rewrite.
  Example:

  “Critique this ceil() implementation for edge cases.”

4. Ask for Error Traps

• Prompt for edge cases and what can break.
  Example:

  “What could break this regex? r'\bT is the (\S+) letter'”

5. Balance Theory with Practice

• Anchor the theoretical with application or testing.
  Example:

  “Given Borges’ Tlön, design a test for LLM ‘independent thought’. Provide pseudocode.”

---

Interaction Style Reflection

• Tone: Direct, efficient.
• Strength: Low fluff, high information density.
• Risk: Might discourage nuance.

Suggested Style Enhancers:

1. “Explain like I’m a beginner.”
2. “Now optimize for brevity.”

---

Optimizing LLM Outputs

1. Pre-constrain Format

“Give a 3-sentence summary, then 3 bullet points of caveats.”

2. Force Prioritization

“Rank these by memory usage for n=1e6.”

3. Meta-Awareness

“What key question did I forget to ask about [topic]?”

---

The BQH Prompt Framework

• Background: Give 1 sentence of context.
• Question: What do you want?
• Hypothesis: Share your assumption (optional).
  Example:

  “SIFT isn’t differentiable — confirm?”

---

Learning and Development Tips

For Programming:

• Automate documentation (docstrings, linters).
• Use test generation (LLM-assisted or libraries).
• Read advanced PyTorch tutorials and profiling tools (e.g., Nsight).

For Theoretical Growth:

• Dive deeper into <insert fields here>.
• Study foundational <insert fields here>.

Research Practices:

• Use reproducibility tools (e.g., DVC).
• Contribute to open-source CV/ML libraries.
• Turn experiments into papers or interactive demos.

---

Using LLMs as a Learning Partner

• Ask for failure cases:

  “What edge cases should I test for this motion vector code?”

• Simulate peer review:

  “Critique this research idea. What assumptions might be invalid?”

• Scaffold interaction:

  “First explain grid_sample, then show how to use it with motion maps.”

---

Final Motto

“Optimize, but leave room for the serendipitous.”

Tweet Share

comments powered by Disqus