System Prompt Analyzer
System Prompt
[sys]
Paste a system prompt to analyze
Detects components, runs quality checks, shows token usage
About
The System Prompt Analyzer reads a system prompt and detects which of 8 components are present: role definition, task description, constraints/rules, output format specification, examples/few-shot, tone/persona, safety rules, and context/background. It also runs 6 quality checks (has role, has format spec, has examples, no obvious contradictions, no excessive repetition, reasonable length) and shows token usage across 4 representative models.
How to use
- 1 Paste a system prompt into the textarea on the left.
- 2 Click "Load example" to see a well-structured system prompt as a reference.
- 3 The right panel shows stats, component detection chips, quality checklist, and token bars.
- 4 Green chips indicate detected components; gray chips indicate absent ones.
- 5 Quality checklist shows ✓ for passing checks and ✗ for failing ones.
- What makes a good system prompt?
- A strong system prompt has: (1) a clear role definition ("You are a senior TypeScript engineer"), (2) a specific task description, (3) explicit constraints ("Never suggest deprecated APIs"), (4) an output format specification ("Respond only with valid JSON"), and (5) at least one example. Role + format spec + one example is the minimum for reliable model behavior.
- How does component detection work?
- Detection uses keyword heuristics: role presence checks for "You are", "Act as", "Your role"; constraints check for "Never", "Always", "Do not", "Must not"; output format checks for "Respond in", "JSON", "Markdown", "Format"; examples check for "Example:", "e.g.", "For instance". This is a fast approximation — complex or unusual system prompts may produce false positives or negatives.
- What is a reasonable system prompt length?
- For most use cases, 200–800 tokens (roughly 150–600 words) is ideal. Very short prompts (<50 tokens) often lack enough specification for reliable behavior. Very long prompts (>4000 tokens) consume significant context window space and can cause the model to lose track of important instructions buried in the middle.