Prompt Templates

All AI prompts in UniPulse follow structured template patterns that ensure consistent, high-quality output from Gemini 2.5 Flash. Prompts are configured via the PROMPT_CONFIGS index and built dynamically with injected context.

Architecture

Prompt Structure Pattern

Every AI call uses a system prompt (defines the AI's role) and a user prompt (provides the specific task):

const response = await callGemini({
  systemPrompt: SYSTEM_PROMPTS.captionGenerator,
  userPrompt: buildCaptionPrompt(input),
  temperature: PROMPT_CONFIGS.caption_generate.temperature,
  maxTokens: PROMPT_CONFIGS.caption_generate.maxTokens,
});

System Prompts

Each AI feature has a dedicated system prompt that defines the AI's role and constraints:

Feature	System Prompt Role
Caption generation	Social media content expert for `{platform}`
Caption rewriting	Content editor specializing in social media optimization
Hashtag generation	Hashtag strategist for `{platform}` algorithm optimization
CTA generation	Marketing copywriter specializing in calls-to-action
Translation	Professional translator for social media content
Conversation reply	Customer support agent for `{brand}`
Intent classification	Message classifier for social media conversations
Content classification	Content categorizer for social media posts
A/B test evaluation	Data analyst specializing in social media experiments
Performance advisor	Social media analytics consultant
Gap analysis	Competitive intelligence analyst

Example: Caption System Prompt

const CAPTION_SYSTEM_PROMPT = `You are a social media content expert.
You generate engaging captions optimized for {platform}.
Follow these guidelines:
- Brand voice: {brandVoice}
- Tone: {tone}
- Language: {language}
- Target audience: {audience}
- Keep within platform character limits
- Use emojis appropriately for the platform
- Never include hashtags in the caption (those are generated separately)`;

Context Injection

Dynamic context is injected into prompts based on available data:

function buildCaptionPrompt(input: CaptionGenerateInput): string {
  const parts = [
    `Topic: ${input.topic}`,
    `Platform: ${input.platform}`,
    `Number of options: ${input.count || 3}`,
  ];

  if (input.brandVoice) {
    parts.push(`Brand voice: ${input.brandVoice.description}`);
    parts.push(`Brand voice samples:\n${input.brandVoice.samples.join('\n')}`);
  }

  if (input.tone) {
    parts.push(`Tone: ${input.tone}`);
  }

  if (input.language && input.language !== 'en') {
    parts.push(`Language: Write in ${input.language}`);
  }

  parts.push(`Generate ${input.count || 3} caption options.`);

  return parts.join('\n');
}

Response Formatting

All prompts request structured JSON output for reliable programmatic parsing:

const prompt = `
${taskDescription}

Respond in the following JSON format:
{
  "captions": [
    {
      "text": "The caption text",
      "tone": "The tone used",
      "characterCount": 150
    }
  ]
}

Important: Return ONLY valid JSON. No markdown, no explanations.
`;

Response Parsing

const response = await callGemini({ systemPrompt, userPrompt, temperature });

// Parse JSON from response
const parsed = JSON.parse(response.text);

// Validate against expected schema
const validated = captionResponseSchema.parse(parsed);