Prompt Power Weekly: Top Tips from the LLM Frontier

Clean prompts, sharper reasoning, better outputs for the latest AI models (including Gemini 3)

Hey friends—another fast, fascinating week across the LLM landscape. I sifted the latest docs, blogs, and research drops and pulled together the prompt patterns that stood out for reliability, clarity, and repeatability. Each tip below includes a copy‑paste template you can try immediately, plus a quick note on why it works. Let’s get you sharper prompts with fewer retries.

---

Top Tips This Week

1) XML‑Tagged “Plan → Execute → Validate” scaffolds (especially strong on Gemini 3)

Summary (source & date): Philipp Schmid shared a concise best‑practice set for Gemini 3: keep instructions precise, place constraints at the top, structure prompts with XML/Markdown tags, and explicitly bridge from long context to your question. It also emphasizes self‑critique and a TODO tracker to keep the model honest (Nov 19, 2025). (Philschmid)

Try this prompt:

<role>You are a senior solutions analyst.</role>
<rules>
1) Be precise and concise.
2) Ask for missing info only if critical.
3) If you assume, flag it explicitly.
</rules>
<planning>
Before answering:
1) Parse goal into sub-tasks.
2) Check if input is sufficient; if not, state what’s missing.
3) Draft a short plan.
</planning>
<context>
[PASTE DATA OR DOCS HERE]
</context>
<task>
Based on the information above, produce:
- Executive Summary (2–3 sentences)
- Bullet Plan (3–6 bullets)
- Open Questions (if any)
</task>
<validate>
Review output against rules and task. Fix any issues before finalizing.
</validate>
<final_instruction>Return only the requested sections.</final_instruction>

Personal insight: The explicit tags reduce “prompt blur” in long contexts, which keeps reasoning tight and outputs consistent—especially when you paste big documents at the top. (Philschmid)