The 80/20 of LLM Prompts
A developer reads a prompt engineering guide, adds six-shot examples, a chain-of-thought reasoning block, output formatting constraints, a system prompt with twelve rules, temperature tuning — and then wonders why the model keeps failing on simple queries.
What happened: The prompt became so complex that it's hard to debug, uses excessive tokens, breaks on uncovered cases, and nobody can maintain it. The model's actual performance barely improved.
What Actually Works: Iterate from Simple
The best prompts are almost boring in their structure:
- Clear task statement — what you want, in one sentence
- Context — only the information the model actually needs
- Output format — how the answer should be structured
- Constraint — what the model should not do
The iteration loop: Write the simplest possible prompt → Test on 20 real cases → Find failure modes → Fix the actual failure → Repeat.
The overlooked variables:
- Temperature: Start at 0.3 for factual tasks, 0.7 for creative tasks.
- Context ordering: Newer and more relevant information should come at the end — the model weights recent content more heavily.
- Token efficiency: If your context window is 90% full with system instructions, there's less room for the actual problem.
Stop using chain-of-thought by default. Use it when: the task has multiple dependent steps, you need to see the model's reasoning, or debugging prompt failures requires it. Don't use it for single-step tasks.
Today's Lesson
80% of prompt improvement comes from 20% of the changes: clarifying the task, removing irrelevant constraints, and testing on real data. Stop over-engineering. Start iterating. Your model's performance will thank you — and so will your token bill.