Llama prompts for landing pages

Every other AI copywriting workflow has the same dependency: you send your copy, your customer data, and your brand strategy to someone else's server. With Llama, you don't have to. Meta's open-source model runs on your hardware, fine-tunes on your data, and generates copy without a single API call leaving your network.

This page isn't about basic "write me a headline" prompts — you can use ChatGPT or Claude for those. This is about building Llama into your landing page workflow: fine-tuning on your brand voice, generating copy in batch, running A/B tests at scale, and analyzing pages without data leaving your servers. These are the prompts and system configurations that make that work.

Set up a brand-voice fine-tuned model

The most powerful thing you can do with Llama is something you can't do with any closed-source model: fine-tune it on your actual copy. After fine-tuning, the model writes in your voice by default — no more "write in a professional but friendly tone" in every prompt.

# Step 1: Prepare your training data
# Collect 50-200 examples of your best copy in JSONL format:

{"messages": [
  {"role": "system", "content": "You are a copywriter for [your brand]. Write in a tone that is [your voice descriptors: direct, technical, slightly irreverent]. Never use [banned words/phrases]. Always [brand rules]."},
  {"role": "user", "content": "Write a landing page headline for our [product feature]"},
  {"role": "assistant", "content": "[Your actual headline that you loved]"}
]}

# Step 2: Fine-tune with your examples
# Using Llama's fine-tuning workflow (via Hugging Face, Axolotl, or your preferred framework):

# Step 3: After fine-tuning, your prompts become much simpler:
Write a landing page headline for [new product/feature].
Context: [one sentence about what it does]
Target audience: [who]

# The model already knows your voice, your word choices, your sentence length preferences.
# No style instructions needed — it writes like you by default.

The training data is everything. Don't use mediocre copy examples — the model learns from what you give it. Curate your 50 best-performing headlines, your tightest benefit blocks, your most compelling CTAs. Quality of examples beats quantity. If you don't have 50 great examples yet, start with the prompts on this page to generate baseline copy, edit it to your voice, then use those edited versions as training data.

Batch landing page copy generation pipeline

If you manage multiple landing pages — across product lines, regions, audiences, or A/B variants — generating copy one page at a time in ChatGPT doesn't scale. This system prompt and batch workflow lets you generate copy for dozens of pages in a single run.

# System prompt for your Llama instance:
You are a conversion-focused copywriter. For each landing page request, generate:
1. HEADLINE (under 10 words, outcome-focused)
2. SUBHEADLINE (1-2 sentences, explains the how)
3. 3 BENEFIT BLOCKS (heading + 2 sentences each)
4. CTA BUTTON TEXT (3-5 words)
5. CTA SUPPORTING LINE (reduces friction, e.g., "No credit card required")
6. META TITLE (under 60 chars, includes primary keyword)
7. META DESCRIPTION (under 155 chars, includes CTA)

Output format: JSON with keys matching the section names above.

# Then feed a batch of page specifications:
---
Page 1:
Product: [product A]
Audience: [audience segment 1]
Keyword: [primary SEO keyword]
Differentiator: [what makes this product/variant unique]
---
Page 2:
Product: [product B]
Audience: [audience segment 2]
Keyword: [primary SEO keyword]
Differentiator: [what makes this product/variant unique]
---
[Repeat for all pages]

# Run as batch inference — no rate limits, no API costs, no per-token pricing.
# Process 50-100 pages in a single run.

The JSON output format matters. When your copy generation outputs structured JSON, you can pipe it directly into your CMS, your build pipeline, or a template system. This is how teams with 100+ landing pages keep copy consistent without a dedicated copywriter on every page. After batch generation, spot-check the output by running sample pages through roast.page to validate quality.

Self-hosted A/B test copy engine

Most A/B testing tools let you create 2-3 variants manually. With a self-hosted Llama instance, you can generate 10-20 variants of any copy element in seconds — and you own the entire pipeline.

# System prompt for A/B variant generation:
You are an A/B testing copy engine. Given a baseline copy element, generate multiple variants that test different psychological approaches. Each variant should change ONE variable so test results are attributable.

Format your output as JSON array with keys: variant_id, copy, hypothesis, variable_changed.

# User prompt:
Baseline: "[your current headline / CTA / subheadline]"
Element type: [headline / CTA / subheadline / benefit block]
Product: [what it does]
Audience: [who they are]

Generate 10 variants. Each should test a different approach:
- 2 variants testing specificity (adding numbers, timeframes)
- 2 variants testing emotional angle (fear vs aspiration)
- 2 variants testing length (shorter vs longer than baseline)
- 2 variants testing social proof integration (weaving in a proof point)
- 2 variants testing direct address (using "you" vs third person)

For each variant, state the hypothesis: "This variant should outperform the baseline because [reason]."

# Example output:
[
  {
    "variant_id": "h_01",
    "copy": "Ship features 3x faster — your backlog will thank you",
    "hypothesis": "Adding a specific multiplier (3x) makes the benefit concrete and testable",
    "variable_changed": "specificity — added quantified outcome"
  },
  ...
]

The structured JSON output means you can feed variants directly into your testing platform's API. No manual copy-paste. And because Llama runs on your hardware, you can regenerate variants as often as needed — daily, weekly, or triggered by test completion. This is the kind of workflow that's cost-prohibitive with per-token API pricing but essentially free with self-hosted Llama.

Privacy-first page analysis

Some companies can't paste their landing page copy into ChatGPT or Claude. Regulated industries, pre-launch stealth products, pages with proprietary pricing — there are legitimate reasons to keep your content off third-party servers. This prompt runs a full page analysis locally.

# System prompt for local page analysis:
You are a senior landing page analyst. Analyze landing page copy for conversion effectiveness. Score each element on a 1-10 scale with specific reasoning. Be critical — high scores should be earned.

# User prompt:
Analyze this landing page copy:

[Paste your full page text — headline through footer CTA]

Product context: [what the product does]
Target audience: [who visits this page]
Conversion goal: [what you want them to do]
Industry: [your industry — helps calibrate expectations]

Evaluate:
1. HEADLINE CLARITY (1-10): Does a stranger understand what this product does in 5 seconds?
2. VALUE PROPOSITION (1-10): Is the "why this over alternatives?" clear?
3. BENEFIT FRAMING (1-10): Are features translated into outcomes?
4. CTA STRENGTH (1-10): Does the CTA tell visitors what they get, not just what to click?
5. TRUST SIGNALS (1-10): Is there enough evidence for a skeptic to believe the claims?
6. COPY FLOW (1-10): Does each section lead logically to the next?
7. OBJECTION HANDLING (1-10): Are likely concerns addressed?

For each score below 7, provide a specific rewrite recommendation.
Overall verdict: Ready to ship / Needs work / Major rewrite needed.

This gives you a private version of what roast.page does with visual analysis. The tradeoff: you lose the visual and layout analysis (Llama processes text, not screenshots), but you gain complete data privacy. For teams under NDA or in stealth mode, this is the pragmatic choice. Use this for copy review, then run the live page through roast.page once it's public for full visual analysis.

API-integrated dynamic copy generation

The most advanced Llama use case for landing pages: generating copy dynamically based on visitor context. This system prompt powers a real-time copy generation API that serves personalized landing page copy.

# System prompt for dynamic copy API:
You are a landing page copy engine integrated via API. Given visitor context, generate personalized copy for specific page sections. Respond in JSON only — no explanations, no markdown, just the JSON object.

# API request format:
{
  "visitor_context": {
    "referrer": "[google-ads / organic / linkedin / direct / email-campaign-name]",
    "industry": "[detected or specified industry]",
    "company_size": "[startup / smb / mid-market / enterprise]",
    "page_variant": "[pricing / features / use-case-specific]"
  },
  "sections_needed": ["headline", "subheadline", "cta_button", "cta_supporting"],
  "product": "[your product, one sentence]",
  "base_value_prop": "[your core value proposition]"
}

# Expected response format:
{
  "headline": "...",
  "subheadline": "...",
  "cta_button": "...",
  "cta_supporting": "...",
  "personalization_notes": "Adjusted for [context detail]"
}

# Example: visitor from Google Ads for "project management tool", enterprise company
# Input context: referrer=google-ads, industry=technology, company_size=enterprise
# Output adapts: headline emphasizes enterprise features, CTA says "Book a Demo" not "Start Free Trial",
# subheadline mentions scale and security rather than ease of use.

This requires engineering to implement — it's not a copy-paste-into-chat workflow. But if your team has the infrastructure, dynamic copy personalization is one of the highest-impact things you can do for conversion rates. Visitors from Google Ads have different intent than organic visitors. Enterprise prospects need different messaging than startups. Self-hosted Llama makes this feasible without per-request API costs eating your margin.

When to use Llama vs closed-source models

• Use Llama when: Data privacy is non-negotiable (regulated industries, pre-launch products, proprietary pricing). When you need batch processing without API costs. When you want to fine-tune on your brand voice. When you're building copy generation into a product or internal tool.
• Use ChatGPT or Claude when: You need the best possible copy quality for a single page. When you don't have GPU infrastructure for self-hosting. When you need it now, today, without setup time. For one-off copy tasks, the quality-to-effort ratio favors closed-source models.
• The hybrid approach: Use ChatGPT or Claude to generate and refine your best copy. Use those outputs as fine-tuning data for Llama. Then use your fine-tuned Llama for all future generation at scale. Best of both worlds.