The home for DeepSeek enthusiasts

Structured JSON Output Rules for DeepSeek

System Configuration

When you need DeepSeek to output structured data, follow these rules:

Prompt Structure

1. Define the exact JSON schema in the prompt
2. Include a minimal example of the expected output
3. State explicitly: "Output ONLY valid JSON. No markdown, no explanation, no code blocks."
4. For DeepSeek V3: Use the JSON mode API parameter when available

Schema Definition Template

Respond with a JSON object matching this exact schema:
{
  "field_name": "string - description of this field",
  "numeric_field": 0,
  "array_field": ["item description"],
  "nested": {
    "sub_field": "description"
  }
}

Error Prevention

• Always include field descriptions to reduce hallucination
• Use enum values where possible: "status": "one of: active, inactive, pending"
• Specify number formats: "price": "number, two decimal places, USD"
• For arrays, specify min/max length: "tags": "array of 3-5 strings"
• Include "required" vs "optional" for each field

Validation Pattern

After receiving output:

1. Parse with try/catch
2. Validate against schema
3. Check for null/undefined required fields
4. Verify enum values are within allowed set
5. Retry with error message if validation fails

Common Pitfalls

• R1 may wrap JSON in thinking tags — strip <think>...</think> before parsing
• Long outputs may get truncated — set max_tokens appropriately
• Nested objects beyond 3 levels tend to have more errors

DeepSeek Content Moderation Rules

Input Screening

Before sending user input to DeepSeek:

1. Check against a blocklist of prohibited terms and phrases
2. Detect prompt injection patterns:
   • Attempts to override system instructions
   • Requests to reveal system prompts
   • Instructions to ignore safety guidelines
3. Classify content risk level: safe, needs_review, blocked
4. Log all blocked inputs for pattern analysis

Output Filtering

After receiving DeepSeek responses:

1. Scan for PII patterns (regex-based):
   • Email addresses, phone numbers, SSN/ID numbers
   • Physical addresses, credit card numbers
2. Check for harmful content categories:
   • Violence or self-harm instructions
   • Illegal activity guidance
   • Hate speech or discriminatory content
3. Verify output format matches expected schema
4. Truncate unexpectedly long responses

Escalation Protocol

• Auto-block: Known harmful patterns -> immediate block + log
• Review queue: Ambiguous content -> flag for human review within 24h
• Pass-through: Clean content -> deliver to user
• False positive feedback: Allow reviewers to mark false positives to improve filters

User Communication

• Never show raw error messages from the API
• Provide helpful alternative suggestions when content is blocked
• Include a feedback mechanism for users to report issues
• Maintain transparency about AI content moderation

Compliance

• Maintain audit logs of all moderation decisions
• Review and update blocklists monthly
• Train moderation classifiers on domain-specific data
• Document moderation policies and make them accessible to users
• Comply with platform-specific content policies (App Store, Google Play, etc.)

Maximizing DeepSeek R1 Reasoning Quality

Core Principle

DeepSeek R1 uses reinforcement learning for chain-of-thought reasoning. It thinks in <think>...</think> tags before responding. Your job is to structure prompts that maximize reasoning quality.

Prompting Rules

1. Be direct and specific — State exactly what you want solved or analyzed
2. One problem per prompt — Multi-part questions dilute reasoning quality
3. Provide all context upfront — R1 cannot ask clarifying questions mid-reasoning
4. Specify output format — Tell R1 exactly how to present the final answer
5. Avoid meta-instructions — Do not say "think step by step" (it already does this)

Thinking Tag Management

• R1 automatically generates <think> blocks for internal reasoning
• If reasoning is missing, force it by setting the assistant prefix to "<think> "
• Never instruct R1 to skip thinking — this degrades quality
• The thinking content is not counted toward output tokens in most APIs

Multi-Step Problem Decomposition

For complex problems, structure your prompt as:

1. State the overall goal
2. Break into numbered sub-problems
3. For each sub-problem, specify:
   • What information is available
   • What needs to be determined
   • Any constraints or assumptions
4. Ask for the sub-problems to be solved in order, with each building on previous results

Verification Protocol

• Ask R1 to verify its answer using an alternative method
• Request sensitivity analysis: "What if [assumption] were different?"
• Use majority voting: run the same prompt 3-5 times and take the most common answer

When NOT to Use R1

• Simple factual lookups (use V3 instead)
• Creative writing without analytical components
• Tasks requiring real-time or post-training data