AI experiments
Sandbox listing LLM prompt structures, token efficiencies, model latencies, and crawler ingestion workers.
LLM benchmarks matrix
| Benchmark Feature | GPT-4o (OpenAI) | Claude 3.5 Sonnet | Gemini 1.5 Flash | Llama 3 (Meta) |
|---|---|---|---|---|
| Primary Strength | Complex reasoning & general coding | Deep architecture & long-context refactoring | Ultra-fast curation & multimodal speed | Privacy, local self-hosted daemons |
| Context Window | 128K tokens | 200K tokens | 1 Million+ tokens | 8K - 128K tokens |
| Latency (Speed) | ~450 ms (Fast) | ~600 ms (Medium) | ~120 ms (Ultra Fast) | Hardware Dependent |
| Cost / 1M Tokens | $2.50 / $10.00 | $3.00 / $15.00 | $0.075 / $0.30 | Self-Hosted ($0 API) |
| JSON Reliability | 99.5% (Strict Mode) | 99.7% (Tool Use) | 99.8% (Zod Validated) | 95.0% (Prompt Constrained) |
| Code Quality Rating | 9.5 / 10 | 9.8 / 10 (Best Architecture) | 9.1 / 10 (Reliable Syntax) | 8.5 / 10 |
Prompt library logs
Deterministic JSON Schema Enforcer
CodingForces LLM APIs to output raw, strictly validated JSON matching Zod schemas without markdown formatting wrappers.
SYSTEM INSTRUCTION:You are a deterministic data transformation pipeline API. You output ONLY valid JSON matching the user schema. Do not output markdown codeblocks, prose, or quotes.
USER PROMPT PAYLOAD:
Act as a structured JSON serializer.
Input Data:
{{input_text}}
Required JSON Output Schema:
{
"title": string,
"summary": string (under 20 words),
"category": "Tech" | "Startup" | "AI",
"confidenceScore": number (0.0 to 1.0),
"tags": string[]
}
Rules:
1. Output ONLY the raw JSON object.
2. Ensure strict key matching and zero trailing commas.Latency: 380ms | Tokens: 420
SEO Article & Metadata Generator
SEOGenerates SEO-friendly tech blogs with headings, metadata, and appropriate JSON-LD schema layouts.
SYSTEM INSTRUCTION:You are an expert tech writer and SEO specialist. Write content that is accurate, factual, readable, and highly optimized for crawlers.
USER PROMPT PAYLOAD:
Act as a senior technology writer. Write a comprehensive guide on the topic: {{topic}}.
Requirements:
1. Include an H1 title and H2/H3 subheadings.
2. Provide a meta description (under 160 characters).
3. Draft the article in clean markdown.
4. Keep the tone professional, educational, and engaging.
5. List 3 key keywords to target.Latency: 1420ms | Tokens: 890
Clean Code & Type-Safety Refactorer
RefactoringRefactors JavaScript/TypeScript code to maximize performance, clean structure, and robust type safety.
SYSTEM INSTRUCTION:You are a principal software engineer. You value type safety, clean code principles, readability, and performance. Do not output explanations unless asked.
USER PROMPT PAYLOAD:
Analyze the following code snippet and refactor it:
```typescript
{{code_snippet}}
```
Refactoring Rules:
- Ensure all types are explicitly defined.
- Optimize loops and asynchronous calls.
- Implement proper error handling.
- Keep helper functions modular.Latency: 1850ms | Tokens: 1240
pgvector Cosine Query Optimizer
CodingOptimizes Supabase & PostgreSQL pgvector similarity queries and HNSW index parameters.
SYSTEM INSTRUCTION:You are a database administrator specializing in PostgreSQL vector embeddings and high-concurrency similarity search indexes.
USER PROMPT PAYLOAD:
Optimize the following pgvector query and index definition: ```sql SELECT id, title, 1 - (embedding <=> $1) AS similarity FROM articles WHERE 1 - (embedding <=> $1) > 0.80 ORDER BY similarity DESC LIMIT 10; ``` Requirements: 1. Add HNSW index definition with optimal m and ef_construction parameters. 2. Tune query execution with SET LOCAL hnsw.ef_search. 3. Explain memory & I/O trade-offs clearly.
Latency: 950ms | Tokens: 780
CTF Log Decoder
DebuggingDecodes hex/base64 representations and performs preliminary security vulnerability checks.
SYSTEM INSTRUCTION:You are a cybersecurity analyst. Help analyze CTF challenge logs without giving direct flags, guiding the learning process.
USER PROMPT PAYLOAD:
Analyze this log snippet:
{{log_text}}
Identify the potential vulnerability category, suggest 3 investigation commands, and describe how to avoid this threat in code.Latency: 980ms | Tokens: 620