I ran 150 identical tasks through ChatGPT (GPT-4o), Claude (Sonnet 4), and Gemini (2.5 Pro). Each task got a raw prompt and a sinc-LLM structured prompt. The biggest finding: the gap between models is smaller than the gap between raw and structured prompts. How you write the prompt matters more than which model you pick.
150 tasks across 5 categories: writing (30), analysis (30), coding (30), research (30), and creative (30). Each task ran raw and structured on all 3 models. That is 900 total runs. I scored each run on accuracy, completeness, format compliance, and constraint adherence.
| Model | Raw Score | Structured Score | Improvement |
|---|---|---|---|
| GPT-4o | 41% | 82% | +100% |
| Claude Sonnet 4 | 46% | 89% | +93% |
| Gemini 2.5 Pro | 39% | 80% | +105% |
The best and worst model on raw prompts differ by only 7 points (46% vs 39%). The gap between raw and structured on the same model is at least 41 points. Prompt structure has 6 times more impact than model selection.
Claude leads on writing with structured prompts (92% vs 85% GPT-4o vs 82% Gemini). Claude follows tone and style rules more closely and sounds more natural. GPT-4o has a recognizable "GPT style" that is hard to turn off. Gemini writes well but sometimes ignores length limits.
Gemini leads on analysis with structured prompts (88% vs 84% Claude vs 81% GPT-4o). Its 1M+ context window lets it handle large documents well. It also reasons through multi-step analysis tasks very well. The DATA band is key here. All three models do much better when you give them real data to work with, instead of asking them to make data up.
Claude leads on coding with structured prompts (86% vs 78% GPT-4o vs 75% Gemini). See the full breakdown in the ChatGPT vs Claude coding comparison. Gemini is a bit behind on coding, but it has improved a lot since early 2025.
Gemini leads on research with structured prompts (87% vs 85% GPT-4o vs 83% Claude). Gemini can search the web. When it does, it finds more current sources. Without web access, all three models are within 3 points of each other.
Claude leads on creative tasks with structured prompts (91% vs 84% GPT-4o vs 78% Gemini). Claude takes more creative risks and follows creative constraints well. Gemini tends to play it safe with creative work.
GPT-4o: Strong PERSONA adherence. It holds its role well across long outputs. Weak on CONSTRAINTS. It tends to "soft-follow" them, meaning it gets close but does not stick to them exactly. Best for tasks where consistency across long outputs matters most.
Claude: Strongest CONSTRAINTS adherence by far. Say "maximum 500 words" and Claude gives you 480 to 500 words. Say "no bullet points" and there are none. Best when precision and rule-following matter more than creative freedom.
Gemini: Strongest DATA processing. It handles large context windows (1M+ tokens) better than the other two. Weaker on FORMAT adherence. It often adds its own structure even when you ask for something specific. Best for analysis of large documents or data-heavy tasks.
The 6-band structure at sinc-LLM works with all three models. It captures universal dimensions of a good prompt, not tricks specific to one model. The same sinc JSON gets high-quality output from GPT-4o, Claude, and Gemini. You can switch models without rewriting your prompts.
Think of it like signal processing. A signal sampled the right way can be read by any decoder. A prompt specified the right way can be read by any LLM. The 6 bands are the Nyquist rate for prompt specification.
| If you need... | Best model | Critical sinc-LLM band |
|---|---|---|
| Precise writing | Claude | PERSONA + CONSTRAINTS |
| Data analysis | Gemini | DATA + FORMAT |
| Coding | Claude | CONSTRAINTS + DATA |
| Research | Gemini | CONTEXT + CONSTRAINTS |
| Creative work | Claude | PERSONA + TASK |
| Quick prototyping | GPT-4o | TASK + FORMAT |
{
"formula": "x(t) = \u03a3 x(nT) \u00b7 sinc((t - nT) / T)",
"T": "specification-axis",
"fragments": [
{"n": 0, "t": "PERSONA", "x": "Expert data scientist with 10 years ML experience"},
{"n": 1, "t": "CONTEXT", "x": "Building a recommendation engine for an e-commerce platform"},
{"n": 2, "t": "DATA", "x": "Dataset: 2M user interactions, 50K products, sparse matrix"},
{"n": 3, "t": "CONSTRAINTS", "x": "Must use collaborative filtering. Latency under 100ms. No PII in logs. Python 3.11+. Must handle cold-start users with content-based fallback"},
{"n": 4, "t": "FORMAT", "x": "Python module with type hints, docstrings, and pytest tests"},
{"n": 5, "t": "TASK", "x": "Implement the recommendation engine with train/predict/evaluate methods"}
]
}
Pick the right model for the job. But always use all 6 bands to structure the prompt. The model matters less than the structure. Start at sincllm.com.
// Production AI Engineering
sinc-LLM designs, audits, and stabilises production AI infrastructure: from vendor evaluation and cost accountability to incident controls and MCP architecture.
See what we do →