🤖 Claude Haiku 4.5 — Smart Rewrite (V4)

Overview

V4 rewrites the existing seo_content from your service JSON using Anthropic's Claude Haiku 4.5 (claude-haiku-4-5-20251001). No web search — with 4 article plan choices (A–D), multi-country locale support, and service registry injection. Distinguishes itself from V2/V3 by exceptional instruction-following and natural prose quality.

Architecture

Flow

┌─────────────────────────────────────────────────────────┐
│  SERVICE DATA (name, category, website, keywords, etc.) │
└────────────────────────┬────────────────────────────────┘
                         │
         ┌──────────────▼──────────────────┐
         │  REWRITE INPUT BLOCK             │
         │  • Existing seo_content (~4K ch)  │
         │  • Service metadata               │
         │  • Cancellation methods           │
         │  • Keywords                       │
         └──────────────┬───────────────────┘
                        │
         ┌──────────────▼───────────────────┐
         │      REWRITE PHASE               │
         │  Rewrite Prompt (Plan A/B/C/D)    │  claude-haiku-4-5-20251001
         │  + system message (brand)         │  max_tokens=16000
         │  + messages [{role:user, ...}]    │  stream=true
         └──────────────┬───────────────────┘
                        │
                  FINAL ARTICLE
                  (HTML, ~1 900 words, 14 H2, 2+ tables)

Heading capitalisation — enforced in system message + QUALITY CHECK self-scan: sentence case on every H1/H2/H3/H4. First word capitalised, proper nouns capitalised, everything else lowercase. Post-generation sanitiser applies as final safety net.

Files involved

• Billoff/functions/api/claude.js — Cloudflare Pages Function proxy (injects ANTHROPIC_API_KEY server-side)
• Billoff/web/assets/openai.js → streamClaude() + generateV4() — Browser-side generator
• Billoff/scripts/04_generate_v4.py — Python batch generator (to create)

Proxy — Cloudflare Pages Function

The Anthropic API key is never exposed to the browser. All calls go through /api/claude:

// functions/api/claude.js — key logic
export async function onRequestPost(context) {
  const apiKey = context.env.ANTHROPIC_API_KEY;     // Cloudflare secret
  const body = await request.text();

  const upstream = await fetch('https://api.anthropic.com/v1/messages', {
    method: 'POST',
    headers: {
      'x-api-key':         apiKey,
      'anthropic-version': '2023-06-01',
      'content-type':      'application/json',
      // Auto-added when body.thinking.type === 'enabled' (used by Analysis):
      // 'anthropic-beta': 'interleaved-thinking-2025-05-14'
    },
    body,
  });
  return new Response(upstream.body, {
    headers: { 'Content-Type': upstream.headers.get('Content-Type') }
  });
}

Client request format

// openai.js — streamClaude()
await fetch('/api/claude', {
  method: 'POST',
  body: JSON.stringify({
    model:      'claude-haiku-4-5-20251001',
    max_tokens: 16000,
    stream:     true,
    system:     systemMsg,
    messages:   [{ role: 'user', content: prompt }],
  })
});

Streaming — Anthropic SSE Format

Claude SSE uses a multi-event protocol. Unlike OpenAI (single data: lines), Anthropic sends pairs of event: + data: lines:

// Stream events (in order)
event: message_start
data: {"type":"message_start","message":{"usage":{"input_tokens":2847,"output_tokens":1}}}

event: content_block_start
data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"<h2>1"}}

// ... more content_block_delta events ...

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":1823}}

event: message_stop
data: {"type":"message_stop"}

The parser in streamClaude() accumulates input_tokens from message_start and output_tokens from message_delta, normalising to {prompt_tokens, completion_tokens} for cost calculation.

Prompt

V4 uses the Rewrite Prompt Template (shared with V2/V3) — an English/AU/Billoff adaptation of the Postclic PROMPT_SEO_CONTENT.txt, plus a dedicated system message:

System message

"Senior consumer advocate at Billoff (billoff.com). "
"'Billoff' MUST appear 5+ times. NEVER 'Postclic'. Active voice. "
"Start with <h1> (creative, non-generic title) then <h2> sections. "
"NEVER write: 'I hope', 'As an AI', 'Based on my knowledge', 'Please note'."
// + WRITING PERSONA block appended (e.g. cancellation_specialist)

Why Claude excels at instruction-following

• Claude 4.x models are specifically trained for precise adherence to formatting instructions
• The system field is processed separately from the user message, giving it higher priority
• Claude rarely hallucinates output format (markdown in place of HTML)
• Word count targets are respected consistently (within ±10%)

Sections 2, 4–14 are identical to V1/V2/V3. Section 3 uses "Cancellation Considerations" (no competitor table — no live data).

🌍 Multi-country & Locale Support

V4 is fully locale-aware. Add locale fields to each service row — get_locale(svc) resolves them and injects country, language, currency and consumer-law references directly into both the system message and the rewrite prompt.

Claude-specific advantage: Claude's strong instruction-following means locale constraints are respected precisely — language mixing and wrong currency symbols are extremely rare.

Output JSON includes cost_usd, cost_eur, elapsed, and stop_reason per article. If stop_reason is max_tokens (truncation), increase MAX_TOKENS or reduce TARGET_WORDS.

Fallback chain: svc.country → editor config → script defaults.

Cost Breakdown

Pricing source: docs.anthropic.com/en/docs/about-claude/pricing (verified Feb 2026)

Model pricing: Input $0.80/1M tokens, Output $4.00/1M tokens (claude-haiku-4-5-20251001).

Prompt caching (available on Claude): if you repeatedly send the same system message + service template, Anthropic charges $0.08/1M for cached reads (vs $0.80/1M). At scale with consistent prompts, this can reduce input costs by up to 90%.

Python Script Reference

Running the batch generator

# From Billoff/ directory
python scripts/04_generate_v4.py

# Compare all 5 methods on 1 service
python scripts/test_compare_3methods.py

Direct API call (Python)

import anthropic

client = anthropic.Anthropic(api_key="sk-ant-...")
message = client.messages.create(
    model="claude-haiku-4-5-20251001",
    max_tokens=8000,
    system="You are a senior SEO writer for Billoff...",
    messages=[{ "role": "user", "content": prompt }],
)
html = message.content[0].text
usage = message.usage   # input_tokens, output_tokens

Environment variables

ANTHROPIC_API_KEY=sk-ant-api03-...
# Cloudflare: set as ANTHROPIC_API_KEY secret in Pages project settings

Pros & Cons

When to use V4

• Pages where prose quality and tone matter more than data freshness
• Large-scale generation (1 000–50 000 articles) where cost needs to stay below V1/V2
• When you want the best "no-search" quality without V1's web search costs
• Projects with highly consistent prompt templates (leverage prompt caching)

All 4 Methods — Head-to-Head

Writing Persona

A writing persona is appended to the V4 system message for every article. Set PERSONA_ID at the top of the downloaded Python script to switch voice. Claude's superior instruction-following means the persona constraints are respected more precisely than any other method. Available: cancellation_specialist (default), consumer_rights_expert, contract_lawyer, financial_advisor. See V1 Docs for the full persona reference.

🔬 AI Comparative Analysis (after all 4 methods)

Once all 4 methods complete, the Lab runs a 2-phase parallel analysis and displays a full comparison report. The analysis is automatically saved to history alongside the article results.

Note: the /api/claude proxy automatically adds the anthropic-beta: interleaved-thinking-2025-05-14 header when the request body contains thinking.type = "enabled". This is required for Claude's Extended Thinking feature used in Phase 2.

Total cost per analysis: ≈€0.04–0.06. See V1 Docs for the full specification.