Occam's Razor for Prompts: Why Your Careful Engineering Limits AI
The Setup
I built an automated research system with Claude Code — daily AI industry news, pushed to Telegram and my blog.
Being a “serious” engineer, I crafted a meticulous prompt:
Search strategy (at least 5 searches, covering different dimensions):
1. US labs: OpenAI, Anthropic, Google DeepMind, Meta, Mistral, xAI
2. China: ByteDance, Baidu, Alibaba, Zhipu, DeepSeek, Moonshot
3. Europe/Japan/Korea
4. High-citation arxiv papers
5. Key figures: Karpathy, Altman, Amodei, LeCun, Sutskever
6. AI policy/regulation/open source
Filtering criteria:
- ✅ Core tech breakthroughs (architecture, training, inference)
- ✅ Major product releases (not minor updates)
- ❌ Marketing fluff, funding news
Output requirements:
- 5-10 items, ranked by importance
- Each item: one sentence what + one sentence why it matters
- Include source links
Format: Telegram HTML...
Each item:
<b>Number. Title</b>
Summary (1-2 sentences)Search dimensions, filtering criteria, output format, item count, per-item structure — all specified. Looks professional. Looks engineered.
Result? 5 brief summaries. One or two sentences each. Barely useful.
The Control
Meanwhile, in Claude App’s Scheduled Tasks, I had this casually written prompt:
As an information analyst, based on your understanding of me, search for the latest
developments in AI, including all major AI labs and new models, all countries' efforts
including the US, Europe, Japan/Korea, China, etc., but avoid meaningless marketing.
I'm an AI practitioner with a strong conviction about AGI, hoping to understand the
latest fundamental industry developments. Also, for important AI figures like Andrej
Karpathy and others, summarize their blogs, articles, etc., and send me all key
AI-related information.No search strategy. No filtering criteria. No format constraints. No item limits.
Result? Eight sections, 50+ deep intelligence items, individual tracking for key figures, trend analysis. Night and day.
Why
I first assumed it was the execution environment — App vs CLI agent loops differ. So I upgraded to a stronger model (Opus), added more search instructions, even added “don’t rush, search thoroughly before summarizing.”
Didn’t help.
Then I replaced the prompt with the same simple version from the App. Same CLI, same model — immediately produced an equally deep report.
The problem was never the execution environment. It was the prompt itself.
What did the carefully designed prompt actually do?
| Constraint | Effect |
|---|---|
| ”5-10 items” | Model actively compresses, discards vast information |
| ”1-2 sentences each” | Deep analysis truncated to summaries |
| ”Ranked by importance” | Model agonizes over ranking instead of coverage |
| ”Search 5 times” | Model treats search as a checklist, not exploration |
| Fixed output template | Model fills forms instead of thinking |
You thought you were “guiding.” You were actually constraining. The model spent its attention budget on following your rules, not on completing the task.
The Insight
This isn’t a prompt technique issue. It’s a mental model issue.
When facing a powerful AI model, what’s your instinct? Most people’s instinct: I need to tell it how to do things. So they write search strategies, filtering criteria, output templates, item limits.
But strong models don’t need you to teach them how to search, filter, or organize. What they need is: who you are, what you want, what matters to you.
❌ "Search 5 times, covering these dimensions..."
✅ "I'm an AI practitioner obsessed with AGI"
❌ "Each item must contain: one sentence + why it matters"
✅ "Tell me the research itself and what it means"
❌ "5-10 items, ranked by importance"
✅ (say nothing, let the model decide)The first is writing an SOP. The second is talking to an expert.
Implications
If this observation holds, a significant portion of “prompt engineering” practice needs reexamination:
- Role + detailed instructions + output template — this classic three-part structure was necessary for weak models, but may be an anti-pattern for strong ones
- “Structured prompt” courses teaching people to write prompts like product requirement documents may be teaching the wrong thing
- Longer prompts are better — this naive intuition is wrong in many scenarios
Occam’s Razor applies to prompts: don’t add unnecessary constraints.
When Do You Need Detailed Prompts?
Not every scenario calls for simple prompts. The distinction:
- Open-ended exploration (search, research, analysis): Simple prompt, let the model judge depth and breadth
- Precise execution (code generation, format conversion, data processing): Detailed prompt, precise output control
The key question: Do you need the model’s judgment, or its execution? When you need judgment, constrain less. When you need execution, constrain more.
One Line
Trust the model’s capability. Tell it who you are, not how to do its job.