Prompt Engineering Conventions

ai-toolsprompt-engineeringllm

Conventions for writing effective prompts, system instructions, and AI interaction patterns in codebases that integrate with LLM APIs like OpenAI, Anthropic, or local models.

Целевые файлы

Файл	Формат
.cursorrules	plaintext
CLAUDE.md	markdown

Содержимое

# Prompt Engineering Conventions ## Prompt Storage - Store all prompts as named constants in a dedicated `prompts/` directory or `lib/prompts.ts` — never inline prompt strings at call sites. - Use template literal functions for dynamic prompts: `const buildSummaryPrompt = (text: string) => \`Summarize: ${text}\``. - Version prompts alongside code. Breaking prompt changes require a version bump and migration note. - Separate system prompts from user message templates — they have different purposes and change at different rates. ## Prompt Structure - System prompts: define role, constraints, output format, and tone. Order matters — put most important constraints first. - Use XML tags to delimit sections in complex prompts: `<context>`, `<instructions>`, `<examples>`, `<output_format>`. - Provide few-shot examples for tasks requiring consistent output format — 2-3 examples outperform lengthy instructions. - State the output format explicitly and early: "Respond with a JSON object matching this schema: ..." - Use negative constraints sparingly — "do not" instructions are less reliable than positive ones. Prefer "return only X" over "do not return Y". ## Dynamic Prompts - Validate all user-supplied inputs before interpolating into prompts — prevent prompt injection. - Escape or sanitize user content: wrap in `<user_input>` tags and instruct the model to treat it as data, not instructions. - Cap injected content length to prevent context overflow: truncate with `...` and note the truncation. - Never interpolate secrets, API keys, or PII into prompts — use placeholder references. ## Model Interaction - Always set `temperature`: 0.0 for deterministic/analytical tasks, 0.3-0.7 for creative tasks, 1.0+ rarely justified. - Always set `max_tokens` to a reasonable limit — never leave it unbounded in production. - Use structured outputs (JSON mode, function calling) whenever the output needs to be parsed programmatically. - Set a timeout for every LLM API call — 30-60 seconds is typical; surface `TimeoutError` to callers. ## Response Parsing - Always parse and validate LLM responses against a schema (Zod, Pydantic) before using the output. - Handle parsing failures gracefully: log the raw response, return a typed error, do not crash. - Never `JSON.parse()` raw LLM output without a try/catch and schema validation. - Implement retry logic with exponential backoff for rate limit errors (429) and transient failures (503). ## Cost & Performance - Log token usage per call in development (`prompt_tokens`, `completion_tokens`, `total_tokens`). - Cache responses for identical prompt+input pairs using a content-addressable key (hash of prompt + input). - Prefer smaller, faster models for classification/routing tasks — reserve large models for generation. - Compress context aggressively: summarize conversation history, chunk documents, use embeddings for retrieval. ## Testing Prompts - Write regression tests with golden outputs for critical prompt + input pairs. - Test prompt edge cases: empty input, very long input, input with special characters, adversarial injection attempts. - Track prompt performance metrics: success rate, average latency, cost per call. - Use deterministic seeds (temperature=0) in tests for reproducible assertions. ## Anti-Patterns - Do not embed business logic inside prompts — keep prompts as pure communication layer. - Do not use prompts as a replacement for traditional code for deterministic operations. - Do not concatenate prompt fragments scattered across multiple files without a central assembly point. - Do not ignore model-specific quirks — what works for GPT-4 may need tuning for Claude or Gemini.

CLI

npx mindaxis apply prompt-engineering-rules --target cursor --scope project

Используется в паках

AI Prompting Toolkit

Master Claude, Codex, and LLM prompt engineering

Открыть пак →

LLM Development Toolkit

Build production AI apps

Открыть пак →

Prompt Engineer Toolkit

AI prompt design and evaluation

Открыть пак →

← Правила

Prompt Engineering Conventions

ai-toolsprompt-engineeringllm

Conventions for writing effective prompts, system instructions, and AI interaction patterns in codebases that integrate with LLM APIs like OpenAI, Anthropic, or local models.

Целевые файлы

Файл	Формат
.cursorrules	plaintext
CLAUDE.md	markdown

Содержимое

CLI

npx mindaxis apply prompt-engineering-rules --target cursor --scope project

Используется в паках

AI Prompting Toolkit

Master Claude, Codex, and LLM prompt engineering

Открыть пак →

LLM Development Toolkit

Build production AI apps

Открыть пак →

Prompt Engineer Toolkit

AI prompt design and evaluation

Открыть пак →